Research Article Segmenting Brain Tissues from …downloads.hindawi.com/journals/bmri/2016/5284586.pdfResearch Article Segmenting Brain Tissues from Chinese Visible Human Dataset by
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Research ArticleSegmenting Brain Tissues from Chinese Visible Human Datasetby Deep-Learned Features with Stacked Autoencoder
1Key Laboratory of Optoelectronic Technology and Systems of Ministry of Education College of Optoelectronic EngineeringChongqing University Chongqing 400044 China2College of Computer and Information Science Chongqing Normal University Chongqing 400050 China3Institute of Digital Medicine College of Biomedical Engineering Third Military Medical University Chongqing 400038 China
Correspondence should be addressed to Xuchu Wang seadriftwanggmailcom
Received 2 November 2015 Revised 18 December 2015 Accepted 27 December 2015
Academic Editor Sher Afzal Khan
Copyright copy 2016 Guangjun Zhao et alThis is an open access article distributed under the Creative CommonsAttribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
Cryosection brain images in Chinese Visible Human (CVH) dataset contain rich anatomical structure information of tissuesbecause of its high resolution (eg 0167mm per pixel) Fast and accurate segmentation of these images into white matter graymatter and cerebrospinal fluid plays a critical role in analyzing andmeasuring the anatomical structures of human brain Howevermost existing automated segmentation methods are designed for computed tomography or magnetic resonance imaging dataand they may not be applicable for cryosection images due to the imaging difference In this paper we propose a supervisedlearning-based CVH brain tissues segmentation method that uses stacked autoencoder (SAE) to automatically learn the deepfeature representations Specifically our model includes two successive parts where two three-layer SAEs take image patches asinput to learn the complex anatomical feature representation and then these features are sent to Softmax classifier for inferring thelabels Experimental results validated the effectiveness of our method and showed that it outperformed four other classical braintissue detection strategies Furthermore we reconstructed three-dimensional surfaces of these tissues which show their potentialin exploring the high-resolution anatomical structures of human brain
1 Introduction
The anatomical structures of the brain tissues are very com-plex and associated with a number of neurological diseasesNevertheless without segmentation the computer cannotrecognize and define a tissuersquos contour automatically andthe anatomical images are difficult to be used for lateralmedical application [1] Cryosection images in the ChineseVisible Human (CVH) dataset show the true color of thehuman body in a high spatial resolution and contain morerich and original details of the brain anatomy than othermedical imaging such as CT and MRI [2] By segmentingCVH brain tissues into cerebrospinal fluid (CSF) graymatter(GM) white matter (WM) or other anatomical structureswe can study human brain and apply it in various fields suchas anatomical education medical image interpretation anddisease diagnosis [3]
It is known that automatic or semiautomatic segmen-tation is helpful for alleviating the laborious and time-consuming manual segment however much noise is intro-duced during CVH image acquisition and the image contrastis low at some positions because of the asymmetric illumi-nation In addition the CVH dataset has no other similardatasets as atlas for guiding segmentation So there remains achallenging problemof how to explore newmodel to segmentthe whole hundreds of CVH brain images in high accuracyand efficiency
Currently most existing brain segmentation algorithmsare based on CT or MRI images According to whether theobjects are labeled these methods can be classified into twocategories unsupervised-based and supervised-based Theunsupervised methods such as region growing threshold-ing clustering and statistical models directly use the imageintensity to search the object For example the fuzzy 119888-means
Hindawi Publishing CorporationBioMed Research InternationalVolume 2016 Article ID 5284586 12 pageshttpdxdoiorg10115520165284586
2 BioMed Research International
method classifies image by grouping similar data that arepresent into clusters and varying the degree of membershipfunction allows the voxel to belong to the multiple classes[4 5]This assumptionmay not work well as it only considersintensity of image and intensity is not enough to expressthe intrinsic feature of objects In addition some methodsestimate distribution of each class with probability density ofGaussian mixture model [6] These methods need accurateestimation of probability density function and for thoseimages with largely overlapped tissues it is hard to matchreal distribution of data in a high accuracy Other methodslike region growing extend threshold by combining it withconnectivity but they need seeds for each region and have theproblem for determining suitable threshold for homogeneityAlso the gradient-based segmentation technique like Water-shed [7] constructs many dams for segmenting image but itis easy to produce oversegmentation
Supervised learning-based segmentation methods arepromising as they take expert information (labeled dataor atlas) into the procedure of segmentation These meth-ods have shown remarkable improvements in segmentingCT or MRI brain images For example Anbeek et al [78] proposed applying spatial and intensity features in apopulation-specific atlas space to label brain voxels Thismethod achieves high accuracy in the cost of a large setof manually segmented training images The studies in [910] proposed using cooccurrence texture features of waveletfollowed by a classifier (like support vector machine) forbrain segmentation while the performance might vary fromdifferent dataset at hand with this kind of low-level featureespecially for the cryosection images that contain many ofanatomical structures and morphological changes
When applying supervised methods to segment theobscure targets there is a common sense that the key tosuccess is mainly dependent on the choice of data representa-tion used to characterize the input data [11] Typical featuressuch as histogram [12] texture [13] and wavelet [10 14]have been successfully applied to many different occasionsBut unfortunately most of these low-level features are hardto extract and organize salient information from the dataand their representation power varied fromdifferent datasetsConsidering that CVH dataset contains hundreds of brainimages with enormous anatomical information of differenttissues due to efficiency we may not label all the imagesas training data but just a tiny fraction so it is crucial toextract better feature representation of the inputs so as toinfer the labels distribution of the unknown anatomical struc-tures
Recently deep neural networks (DNN) have shown theirpromising results for feature extraction in many computervision applications [15 16] Contrary to traditional shallowclassifiers in which feature engineering is crucial deeplearning methods automatically learn hierarchies of rele-vant features directly from the raw inputs [17] There areseveral DNN-based models such as convolutional neuralnetworks (CNNs) [16] deep belief network (DBN) [18]stacked autoencoder (SAE) [19ndash21] that have been applied indifferent tasks As a typical example of deep models CNNsalternatingly apply trainable filters and local neighborhood
pooling operations on the raw input images resulting in ahierarchy of increasingly complex features [22 23]
The first deep autoencoder network was proposed byHinton and Salakhutdinov in [20] In contrast to CNNsthat apply a series of convolutionpoolingsubsampling oper-ations to learn deep feature representations SAE employsa full connection of units for deep feature learning SAEcontains multiple intermediate layers and millions of train-able parameters that enable it to capture highly nonlinearmapping between input and output so recently it has beenwidely applied in some tasks such as image denoising anddeconvolution [24 25] multiple organ detection [26] infanthippocampus segmentation [19] and nuclei regions extrac-tion [27]Their existing results indicate that the archtecture ofSAE is essential for acheiving better performance in a specfictask which motivates us to investigate the SAE-based featurelearning for CVH brain segmentation
In this paper we propose a learning-based CVH braintissues segmentation model that employs unsupervised SAEto automatically learn the deep feature representations andsupervised Softmax for classificationOurmodel is composedof two successive parts white matter (WM) segmentationmodule and gray matter (GM)cerebrospinal fluid (CSF)segmentation module Specifically the SAE in two modulestakes image patches as input and learns their deep featurerepresentation These features are then sent to a Softmaxclassifier for inferring the labels of the center pixels of thesepatches To decrease the burden of labeling only a tinynumber of labeled anatomical patches are fed to the networkin the training stage Intuitively a trained model may bestrange to the new patches that contain unknown anatomicalstructures but because SAE can learn intrinsic feature repre-sentations which are well in eliminating distortion rotationand blurring of the input patches the model can infer theclasses of patches that contain unknown anatomical struc-tures The proposed model was used to segment all 422 CVHbrain images at hand And the segmentation performanceof the deep-learned feature was compared with some otherrepresentative features (eg intensity PCA HOG and AE)Experimental results show that the proposed model achieveshigher accuracy in segmenting all three tissues
The rest of the paper is organized as follows Section 2briefly reviews the acquisition of CVH dataset and the detailsof the proposed model Section 3 reports the experimentalresults and analyzes the segmentation performance of dif-ferent SAE architecture It also compared the performanceof different methods and visualizes the three-dimensionalreconstruction results At last Section 4 concludes the paper
2 Material and Methods
21 Image Acquisition and Preprocessing The data utilizedin this study are successive cross-sectional images of humanbrain from the CVH dataset provided by the Third MilitaryMedical University in ChinaThe cadaver is 162 cm in height54 kg in weight and free of organic lesions Both the donorand her relatives donated their bodies to the Chinese VisibleHuman program which follows scientific ethics rules of theChinese Ethics Department
BioMed Research International 3
(a) (b) (c)
Figure 1 Preprocessing example of cryosection brain image (a) Original imagewithout any preprocessing (3072times 2048 pixels) (b) Croppedimage (1252 times 1364 pixels) (c) Skull stripped image
SAE
Input patches Input patches
Softmax
3-hidden-layer SAE
Softmax
MODEL 1 MODEL 2
SAE
Ground truth Ground truth
3-hidden-layer SAE
(120596 times 120596 times 120598) (120596 times 120596 times 120598)W1 W2
W3
W4
W1 W2
W3 W4
minus
Training MODEL 2 towards GM and CSFsegmentation
Application of segmentation
Training MODEL 1 towards WM segmentation
Figure 2 Flowchart of our segmentation model
The images in CVH dataset are taken of the frozencadaver A total of 422 cross-sectional images of the head(number 1074 to number 1495) are selected for this study Asshown in Figure 1 the slice is 0167mm per pixel 025mmthick and photographed at a resolution of 6291456 (3072times 2048) pixels with 24-bit color information in tiff format[28] In order to reduce computational cost and memoryusage these images are transformed into PNG format andcropped to 1252 times 1364 pixels In the preprocessing stageskull stripping is applied to each image
22 Method Overview In this work the CVH brain tissuesegmentation problem is formulated as a patch classificationtask and the architecture of our segmentationmodel is shownin Figure 2 The model takes patches extracted from the B-channel and V-channel of the original images as input thenSAE is used to extract intrinsic feature representation of the
input patches and the following Softmax classifier generates alabels distribution of these patches based on the deep featuresThis model on segmenting three brain tissues (CSF GMand WM) actually contains two submodels MODEL 1 andMODEL 2 In MODEL 1 GM and CSF are labeled as thesame class and the segmentation is formulated as a three-class classification task WM GM amp CSF and backgroundThrough MODEL 1 WM tissue can be extracted from theregion of interest This segmentation result is helpful to fillthe areas of WM into background so as to eliminate theinfluence of WM So in MODEL 2 the patches from theWM-eliminated image are taken as inputs and the image issegmented into CSF GM and background This pipeline hasthe advantages in that more image patches of the objects withfewer labeled data can be taken as it is quite time-consumingto manually label an image In the following we will describethe details of the model
4 BioMed Research International
W1
W2
W3
EncoderDecoder
h1
n h2
n
h3
n
xn
Figure 3 Proposed three-hidden-layer SAE Note that the number of layers in our model is set via cross-validation
23 Learning Hierarchical Feature Representation by SAE
231 Single-Layer AE A single-layer AE [29] is a kind ofunsupervised neural network whose goal is to minimizethe reconstruction error from inputs to outputs via twocomponents encoder and decoder [19] In the encodingstage given an input sample
119899isin R119873 AE will map it to the
hidden activation ℎ119899isin R119872 by the following mapping
ℎ119899= 119891 (W
1119899+1198871) (1)
where 119891(119911) = 1(1 + exp(minus119911)) is a nonlinear activationfunction W
1isin R119872times119873 is the encoder weight matrix 119887
1isin
R119872 is the bias vector Generally119872 lt 119873 then the networkis forced to learn a compressed representation of the inputvector This compressed representation can be viewed asfeatures of the input vector In the decoding stage AE willreproduce input data from the hidden activation ℎ
119899isin R119872 by
119910119899= 119891 (W
2
ℎ119899+1198872) (2)
whereW2isin R119873times119872 is the decode weight matrix and 119887
2isin R119873
is the bias vectorIn our model during the training stage we minimize the
objective function shown in (3) with respect to the weightmatrixesW
1andW
2and bias vectors 119887
1and 1198872 The objective
function includes an average sum-of-square error term tofit the input data and a weight decay term to decrease themagnitude of weight matrices as well as helping prevent over-fitting
119869 (W b) = 1119898
119898
sum
119899=1
(
1
2
1003817100381710038171003817x119899minus y119899
1003817100381710038171003817
2
)
+
120582
2
119899119897minus1
sum
119897=1
119904119897
sum
119894=1
119904119897+1
sum
119895=1
(119882(119897)
119895119894)
2
(3)
where x119899denotes the 119899-th sample in the training set y
119899
denotes the reconstructed output with input of x119899 120582 denotes
weight decay parameter which controls the relative impor-tance of the two terms 119899
119897denotes the number of layers in the
network 119904119897denotes the number of units in the 119897th layer and
119898 denotes the number of training samples
232 SAE for Hierarchical Feature Learning SAE is a typeof neural network consisting of multiple layers of AEs inwhich the output of each layer is wired to the inputs of thesuccessive layer In this paper we propose a multi-hiddenlayer SAE which is shown in Figure 3 It is noted that thenumber of layers in our model is set via cross-validationFor an input vector
119899 the first layer transforms it into a
vector ℎ1119899that consists of activations of hidden units and the
second layer takes ℎ1119899as input to produce a new activated
vector ℎ2119899 then the final activated vector ℎ3
119899that is produced
by ℎ2119899can be viewed as deep-learned feature representation
of the input sample It is noticed that the model intrinsicallyhandles varying-dimension images through image patcheswith different sizes For a specific task the parameters areusually settled through experiments or experience and thetraining and application of SAE will go on those parameters
In our task we follow the greedy layer-wise trainingstrategy [18 20 30] to obtain better parameters of a SAEThatis we first train the first single AE on the raw input and thentrain the second AE on the hidden activation vector acquiredby the former AE The subsequent layers are repeated usingthe output of each layer as input Once this phase of trainingis complete we stack AEs into SAE and train the entirenetwork by a gradient-based optimization method to refinethe parameters
The high-level features learned by SAE are more discrim-inative compared to hand-crafted feature such as intensityand learning-based feature by single-layer AE To make anintuitive interpretation we conducted a dimension reductionexperiment to visually examine the distributions of featurevectors from image patches by original intensity and a SAEwith three hidden layers respectivelyThe experimental resultis shown in Figure 4 where the dimensionality of each featurevector is reduced to two by Principal Components Analysis(PCA) for the purpose of visualization We can see thatthe features extracted by SAE output a better cluster resultthan intensity features It is easier to generate a separationhyperplane for separating different types of samples
To further visualize the discriminative ability betweenAE and SAE the features learned by each layer of SAEbased on cryosection image are shown in Figure 5 As shownin Figure 5(a) it is seen that AE can only learn primitiveoriented edge-like features just like119870-means ICA or sparse
BioMed Research International 5
WMGMCSF
0 10 20 30minus20 minus10minus30minus15
minus10
minus5
0
5
10
15
(a)
WMGMCSF
0 1 2 3 4minus2minus3minus4 minus1minus5minus2
minus15
minus1
minus05
0
05
1
15
2
25
(b)
Figure 4 Two-dimensional feature representation for 500 patches of each brain tissue by (a) intensity + PCA and (b) intensity + three-hidden-layer SAE + PCA here PCA is just for visualization of principle components
(a) (b) (c)
Figure 5 Visualization of learned high-level features of input pixel intensities with three-layer SAE (a) (b) and (c) Learned featurerepresentation in the first (with 225 units) second (with 144 units) and third (with 81 units) hidden layers respectively where the features inthe third layer are discriminative for image segmentation task
coding do [31] While SAE can learn higher-level featurescorresponding to the patterns in the appearance of featuresin the former layer (as shown in Figure 5(c)) these high-levelfeatures are more discriminative for image segmentation taskin this paper Hence our model employs the SAE insteadof both hand-craft features and learning based features bysinger-layer AE to extract high-level feature representationfor segmenting brain tissues
233 SAE Plus Softmax for CVH Brain Tissues SegmentationFor every foreground pixel in the cryosection image weextract two patches centered at this pixel from its B-channel(in RGB color space) and V-channel (in HSV color space)image respectively So 120598 in Figure 2 is 2 The two patchesare concatenated together as the input features of SAE andthe features learned by SAE then are sent to a supervised
Softmax classifier The parameters of two SAEs in MODEL1 and MODEL 2 are roughly consistent The patch sizeis set via cross-validation and layer depth is set among1 2 3 considering the balance between computation costand discriminative power The number of units in each layerof SAE is experimentally set as 400 200 and 100 respectivelyThus the final dimensionality of deep-learned feature is 100The weight decay 120582 of two SAEs in MODEL 1 and MODEL 2is set to 0003 and 0005 respectively which is tuned on thevalidation set
234 Ground Truth and Training Sets Generation Theobjec-tive of the proposed model is to automatically segment thethree brain tissues of thewhole 422CVHbrain image at handIt is laborious and time-consuming to manually segmentall the brain image for an anatomical expert so the expert
6 BioMed Research International
Table 1 Details of the SAE architectures with different input patch size and layer depth in this study
only segmented eight CVH brain images which is saturatedwith abundant anatomical structures as the ground truth fortraining sets generation and quantitative evaluation
The patches used for training and testing are extractedfrom the eight labeled images The size of each patch ischosen 17 times 17 to achieve relatively best performance whichis validated in Section 31 so 120596 in Figure 2 is 17 For MODEL1 our training sets contain 10000WM patches and 90000non-WM patches extracted from eight training images Formodel 2 the training sets contain 106000GM patches and84000CSF patches These patches are used for SAE andSoftmax training in the two models
3 Experimental Results and Discussion
In the experiments we firstly focus on evaluating the seg-mentation performance of features learnt by different SAEsbased on cryosection image in order to get the best SAEarchitecture Secondly we compare performances of theproposedmodel based on the deep-learned features and somefamous hand-crafted features such as intensity PCA andHOGand one learning-based feature byAEThen we presenttypical segmentation examples and make some discussionFinally we build the 3D meshes of three tissues based on oursegmentation results
31 Comparison ofDifferent SAEArchitectures Thenonlinearmapping between the input and output of SAE is influencedby its multilayer architecture with various input patch sizesand depths In order to investigate the impact of different SAEarchitectures on segmentation accuracy five different SAEarchitectures are designed and resort to segmentation taskThe detailed parameter configurations are shown in Table 1and the segmentation performances of WM GM and CSFare reported in Table 2
It can be observed from the results that the predictiveaccuracies are generally higher for the architectures withinput patch sizes of 17 times 17 and 21 times 21 The SAEs with
Table 2 Mean and standard deviation of Dice ratio (in ) formeasuring the performance of the three tissue types with fivedifferent architectures trained by using different patch sizes of 5 times 59 times 9 13 times 13 17 times 17 and 21 times 21 respectively The experimentswere conducted in a leave-one-out manner and eight test resultswere collected for each tissue
larger patch size tend to have a deeper hierarchical structureand more trainable parameters These learned parametersare capable of capturing the complex relationship betweeninput and output We can also observe that the architecturewith input patch size of 21 times 21 does not generate substan-tially higher performance suggesting that larger patch mayintroducemore image noise and fused anatomical structuresIn order to obtain better segmentation performance in thefollowing we focus on evaluating the performance of our SAEarchitecture with input patch size of 17 times 17 and depth of 3
32 Comparison of Performances Based on Different Fea-tures In order to provide a comprehensive evaluation ofthe proposed method and illustrate the effect of high-levelfeatures in contrast to low-level features three representativehand-crafted features such as intensity PCA [32 33] andHOG [34] and one learning-based features by AE are usedfor comparison These features follow the same segmentingprocedures as the deep-learned features All the segmentationperformances are reported in Table 3 using Dice ratio It canbe observed that our model with the features extracted bySAE outperforms other well-known features for segmentingall three types of brain tissue Specifically our model yields
BioMed Research International 7
(a)
(b)
(c)
(d)
(e)
Figure 6 Continued
8 BioMed Research International
(f)
(g)
Figure 6 Comparison of the segmentation results with the manually generated segmentation on a cryosection image in CVH dataset (a)shows the original cryosection and its B-channel (in RGB color space) and V-channel (in HSV color space) image (b) shows the manualresults (CSF GM and WM) (c)ndash(g) show the results by the features of our SAE deep learning Intensity PCA HOG and AE respectively
Table 3 Mean and standard deviation of Dice ratio (in ) for thesegmentations obtained by five feature representation methods
the best average Dice ratio of 9069 plusmn 214 (CSF) 9124plusmn 201 (GM) and 9612 plusmn 123 (WM) in a leave-one-outevaluation manner These results have illustrated the strongdiscriminative power of the deep-learned features in braintissues segmentation task
To further demonstrate the advantages of our proposedmodel we visually examine the segmentation results on onecryosection image which is shown in Figure 6 (a) shows theoriginal RGB cryosection image and its B- and V-channelimagesThe ground truth that segmented by experts is shownin (b) (c)ndash(g) present segmentation results of four methodsbased on deep-learned features intensity features PCAfeatures HOG features and AE feautres respectivelyWe cansee that the segmentation results of the proposed model arequite close to the ground truth In contrast other resultseither generate much oversegmentation or fail to segmenttiny anatomical structures accurately Specifically WM isonly adjacent to the GM and can be easily distinguished
from its surroundings so the WM segmentation DRs areapproximate to each other And the visible results of WMare similar in appearance except the fact that HOG-basedmethod mistakenly introduced a small fraction of CSFinto the results It has some difficulty on GM and CSFsegmentation due to the complex anatomical structures andlow contrast HOG and intensity based methods introducemore surroundingnon-GMornon-CSF tissue into theROI ofGM or CSF respectively thus they produce more defects andfuzzy boundaries for different tissues In contrast theGMandCSF tissues generated by ourmethod can be clearly identifiedwith a certain ROI and distinct contours
We then applied our proposed model to segment all422 brain cryosection images Typical images in coronal andsagittal viewpoints and their corresponding segmentationresults (WM GM and CSF) are shown in Figures 7 and 8respectively It is remarkably seen that the results of WMand GM change continuously and their morphological dis-tributions are shown clearly most tiny anatomical structuresare well reserved It is also noticed that the results of CSFseem incomplete and not distributed uniformly This fact isdetermined by the characteristics of the cryosection imagesThese images have such high spatial resolution (0167mmper pixel and 025mm per slice) that it can express finestructures of tissues But since these images were collectedfrom a cadaver the CSF in the brain no longer flowed invivo Due to the effect of gravity the CSF will gather to someplaces rather than uniformly distributing around the surfaceof brain as in live statusThese factors cause the discontinuousdistribution of the CSF segmentation results
BioMed Research International 9
(a)
(b)
(c)
(d)
Figure 7 Coronal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
(a)
(b)
(c)
(d)
Figure 8 Sagittal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
10 BioMed Research International
Figure 9 Three-dimensional surface-rendering reconstruction results of WM GM and CSF based on our segmented images
33 Three-Dimensional Reconstruction Results The CVHcryosection slices are 0167mm per pixel and 025mm thickand such high resolution is very helpful for displaying subtleanatomical structures of brain tissues For a more in-depthunderstanding of these tissues the segmented white mattergray matter and cerebrospinal fluid images are reconstructedusingmarching cubes algorithm to produce 3D surfacemeshThe reconstruction results in different views are shown inFigure 9 From the surface-rendering reconstruction resultsit is seen that the surface of WM is smooth and its sulci andfissures are clearly displayed The distribution of GM shapeis also noticed but the surface of GM reconstruction resultsdoes not look very smooth The reason for this lies in thefact that the GM and CSF are mixed together because of theice crystals in the frozen brain slices so the segmentation ofGM is influenced by its surrounding CSF In spite of it thesulci and cerebral cisterns are also easy to be recognized in3D reconstructed WM and GM
Benefitting from the increasing development of the 3Dreconstruction technology 3D MRI and PET have nowbeen used in clinic and researches But because of theresolution limitation and complexity of brain structures ofthe 2D radiological images (such as CT and MRI) the 3Dreconstruction results are usually unsatisfactory and are hardfor the guidance of clinic operation For the work in ourpaper we focus on the segmentation and reconstruction ofthree kinds of CVHbrain tissuesTheCVHbrain images havehigh spatial resolution of 0167mmper pixel and 025mmperslice after segmentation by the proposed method the high-quality and high-accuracy 2D brain tissues can get well 3Dreconstruction results Some potential applications of the 3Dreconstruction result include the following
(1) The 3D results can be viewed in any orientationbesides the common coronal sagittal and transverseorientations These 3D models (especially for WM
BioMed Research International 11
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
method classifies image by grouping similar data that arepresent into clusters and varying the degree of membershipfunction allows the voxel to belong to the multiple classes[4 5]This assumptionmay not work well as it only considersintensity of image and intensity is not enough to expressthe intrinsic feature of objects In addition some methodsestimate distribution of each class with probability density ofGaussian mixture model [6] These methods need accurateestimation of probability density function and for thoseimages with largely overlapped tissues it is hard to matchreal distribution of data in a high accuracy Other methodslike region growing extend threshold by combining it withconnectivity but they need seeds for each region and have theproblem for determining suitable threshold for homogeneityAlso the gradient-based segmentation technique like Water-shed [7] constructs many dams for segmenting image but itis easy to produce oversegmentation
Supervised learning-based segmentation methods arepromising as they take expert information (labeled dataor atlas) into the procedure of segmentation These meth-ods have shown remarkable improvements in segmentingCT or MRI brain images For example Anbeek et al [78] proposed applying spatial and intensity features in apopulation-specific atlas space to label brain voxels Thismethod achieves high accuracy in the cost of a large setof manually segmented training images The studies in [910] proposed using cooccurrence texture features of waveletfollowed by a classifier (like support vector machine) forbrain segmentation while the performance might vary fromdifferent dataset at hand with this kind of low-level featureespecially for the cryosection images that contain many ofanatomical structures and morphological changes
When applying supervised methods to segment theobscure targets there is a common sense that the key tosuccess is mainly dependent on the choice of data representa-tion used to characterize the input data [11] Typical featuressuch as histogram [12] texture [13] and wavelet [10 14]have been successfully applied to many different occasionsBut unfortunately most of these low-level features are hardto extract and organize salient information from the dataand their representation power varied fromdifferent datasetsConsidering that CVH dataset contains hundreds of brainimages with enormous anatomical information of differenttissues due to efficiency we may not label all the imagesas training data but just a tiny fraction so it is crucial toextract better feature representation of the inputs so as toinfer the labels distribution of the unknown anatomical struc-tures
Recently deep neural networks (DNN) have shown theirpromising results for feature extraction in many computervision applications [15 16] Contrary to traditional shallowclassifiers in which feature engineering is crucial deeplearning methods automatically learn hierarchies of rele-vant features directly from the raw inputs [17] There areseveral DNN-based models such as convolutional neuralnetworks (CNNs) [16] deep belief network (DBN) [18]stacked autoencoder (SAE) [19ndash21] that have been applied indifferent tasks As a typical example of deep models CNNsalternatingly apply trainable filters and local neighborhood
pooling operations on the raw input images resulting in ahierarchy of increasingly complex features [22 23]
The first deep autoencoder network was proposed byHinton and Salakhutdinov in [20] In contrast to CNNsthat apply a series of convolutionpoolingsubsampling oper-ations to learn deep feature representations SAE employsa full connection of units for deep feature learning SAEcontains multiple intermediate layers and millions of train-able parameters that enable it to capture highly nonlinearmapping between input and output so recently it has beenwidely applied in some tasks such as image denoising anddeconvolution [24 25] multiple organ detection [26] infanthippocampus segmentation [19] and nuclei regions extrac-tion [27]Their existing results indicate that the archtecture ofSAE is essential for acheiving better performance in a specfictask which motivates us to investigate the SAE-based featurelearning for CVH brain segmentation
In this paper we propose a learning-based CVH braintissues segmentation model that employs unsupervised SAEto automatically learn the deep feature representations andsupervised Softmax for classificationOurmodel is composedof two successive parts white matter (WM) segmentationmodule and gray matter (GM)cerebrospinal fluid (CSF)segmentation module Specifically the SAE in two modulestakes image patches as input and learns their deep featurerepresentation These features are then sent to a Softmaxclassifier for inferring the labels of the center pixels of thesepatches To decrease the burden of labeling only a tinynumber of labeled anatomical patches are fed to the networkin the training stage Intuitively a trained model may bestrange to the new patches that contain unknown anatomicalstructures but because SAE can learn intrinsic feature repre-sentations which are well in eliminating distortion rotationand blurring of the input patches the model can infer theclasses of patches that contain unknown anatomical struc-tures The proposed model was used to segment all 422 CVHbrain images at hand And the segmentation performanceof the deep-learned feature was compared with some otherrepresentative features (eg intensity PCA HOG and AE)Experimental results show that the proposed model achieveshigher accuracy in segmenting all three tissues
The rest of the paper is organized as follows Section 2briefly reviews the acquisition of CVH dataset and the detailsof the proposed model Section 3 reports the experimentalresults and analyzes the segmentation performance of dif-ferent SAE architecture It also compared the performanceof different methods and visualizes the three-dimensionalreconstruction results At last Section 4 concludes the paper
2 Material and Methods
21 Image Acquisition and Preprocessing The data utilizedin this study are successive cross-sectional images of humanbrain from the CVH dataset provided by the Third MilitaryMedical University in ChinaThe cadaver is 162 cm in height54 kg in weight and free of organic lesions Both the donorand her relatives donated their bodies to the Chinese VisibleHuman program which follows scientific ethics rules of theChinese Ethics Department
BioMed Research International 3
(a) (b) (c)
Figure 1 Preprocessing example of cryosection brain image (a) Original imagewithout any preprocessing (3072times 2048 pixels) (b) Croppedimage (1252 times 1364 pixels) (c) Skull stripped image
SAE
Input patches Input patches
Softmax
3-hidden-layer SAE
Softmax
MODEL 1 MODEL 2
SAE
Ground truth Ground truth
3-hidden-layer SAE
(120596 times 120596 times 120598) (120596 times 120596 times 120598)W1 W2
W3
W4
W1 W2
W3 W4
minus
Training MODEL 2 towards GM and CSFsegmentation
Application of segmentation
Training MODEL 1 towards WM segmentation
Figure 2 Flowchart of our segmentation model
The images in CVH dataset are taken of the frozencadaver A total of 422 cross-sectional images of the head(number 1074 to number 1495) are selected for this study Asshown in Figure 1 the slice is 0167mm per pixel 025mmthick and photographed at a resolution of 6291456 (3072times 2048) pixels with 24-bit color information in tiff format[28] In order to reduce computational cost and memoryusage these images are transformed into PNG format andcropped to 1252 times 1364 pixels In the preprocessing stageskull stripping is applied to each image
22 Method Overview In this work the CVH brain tissuesegmentation problem is formulated as a patch classificationtask and the architecture of our segmentationmodel is shownin Figure 2 The model takes patches extracted from the B-channel and V-channel of the original images as input thenSAE is used to extract intrinsic feature representation of the
input patches and the following Softmax classifier generates alabels distribution of these patches based on the deep featuresThis model on segmenting three brain tissues (CSF GMand WM) actually contains two submodels MODEL 1 andMODEL 2 In MODEL 1 GM and CSF are labeled as thesame class and the segmentation is formulated as a three-class classification task WM GM amp CSF and backgroundThrough MODEL 1 WM tissue can be extracted from theregion of interest This segmentation result is helpful to fillthe areas of WM into background so as to eliminate theinfluence of WM So in MODEL 2 the patches from theWM-eliminated image are taken as inputs and the image issegmented into CSF GM and background This pipeline hasthe advantages in that more image patches of the objects withfewer labeled data can be taken as it is quite time-consumingto manually label an image In the following we will describethe details of the model
4 BioMed Research International
W1
W2
W3
EncoderDecoder
h1
n h2
n
h3
n
xn
Figure 3 Proposed three-hidden-layer SAE Note that the number of layers in our model is set via cross-validation
23 Learning Hierarchical Feature Representation by SAE
231 Single-Layer AE A single-layer AE [29] is a kind ofunsupervised neural network whose goal is to minimizethe reconstruction error from inputs to outputs via twocomponents encoder and decoder [19] In the encodingstage given an input sample
119899isin R119873 AE will map it to the
hidden activation ℎ119899isin R119872 by the following mapping
ℎ119899= 119891 (W
1119899+1198871) (1)
where 119891(119911) = 1(1 + exp(minus119911)) is a nonlinear activationfunction W
1isin R119872times119873 is the encoder weight matrix 119887
1isin
R119872 is the bias vector Generally119872 lt 119873 then the networkis forced to learn a compressed representation of the inputvector This compressed representation can be viewed asfeatures of the input vector In the decoding stage AE willreproduce input data from the hidden activation ℎ
119899isin R119872 by
119910119899= 119891 (W
2
ℎ119899+1198872) (2)
whereW2isin R119873times119872 is the decode weight matrix and 119887
2isin R119873
is the bias vectorIn our model during the training stage we minimize the
objective function shown in (3) with respect to the weightmatrixesW
1andW
2and bias vectors 119887
1and 1198872 The objective
function includes an average sum-of-square error term tofit the input data and a weight decay term to decrease themagnitude of weight matrices as well as helping prevent over-fitting
119869 (W b) = 1119898
119898
sum
119899=1
(
1
2
1003817100381710038171003817x119899minus y119899
1003817100381710038171003817
2
)
+
120582
2
119899119897minus1
sum
119897=1
119904119897
sum
119894=1
119904119897+1
sum
119895=1
(119882(119897)
119895119894)
2
(3)
where x119899denotes the 119899-th sample in the training set y
119899
denotes the reconstructed output with input of x119899 120582 denotes
weight decay parameter which controls the relative impor-tance of the two terms 119899
119897denotes the number of layers in the
network 119904119897denotes the number of units in the 119897th layer and
119898 denotes the number of training samples
232 SAE for Hierarchical Feature Learning SAE is a typeof neural network consisting of multiple layers of AEs inwhich the output of each layer is wired to the inputs of thesuccessive layer In this paper we propose a multi-hiddenlayer SAE which is shown in Figure 3 It is noted that thenumber of layers in our model is set via cross-validationFor an input vector
119899 the first layer transforms it into a
vector ℎ1119899that consists of activations of hidden units and the
second layer takes ℎ1119899as input to produce a new activated
vector ℎ2119899 then the final activated vector ℎ3
119899that is produced
by ℎ2119899can be viewed as deep-learned feature representation
of the input sample It is noticed that the model intrinsicallyhandles varying-dimension images through image patcheswith different sizes For a specific task the parameters areusually settled through experiments or experience and thetraining and application of SAE will go on those parameters
In our task we follow the greedy layer-wise trainingstrategy [18 20 30] to obtain better parameters of a SAEThatis we first train the first single AE on the raw input and thentrain the second AE on the hidden activation vector acquiredby the former AE The subsequent layers are repeated usingthe output of each layer as input Once this phase of trainingis complete we stack AEs into SAE and train the entirenetwork by a gradient-based optimization method to refinethe parameters
The high-level features learned by SAE are more discrim-inative compared to hand-crafted feature such as intensityand learning-based feature by single-layer AE To make anintuitive interpretation we conducted a dimension reductionexperiment to visually examine the distributions of featurevectors from image patches by original intensity and a SAEwith three hidden layers respectivelyThe experimental resultis shown in Figure 4 where the dimensionality of each featurevector is reduced to two by Principal Components Analysis(PCA) for the purpose of visualization We can see thatthe features extracted by SAE output a better cluster resultthan intensity features It is easier to generate a separationhyperplane for separating different types of samples
To further visualize the discriminative ability betweenAE and SAE the features learned by each layer of SAEbased on cryosection image are shown in Figure 5 As shownin Figure 5(a) it is seen that AE can only learn primitiveoriented edge-like features just like119870-means ICA or sparse
BioMed Research International 5
WMGMCSF
0 10 20 30minus20 minus10minus30minus15
minus10
minus5
0
5
10
15
(a)
WMGMCSF
0 1 2 3 4minus2minus3minus4 minus1minus5minus2
minus15
minus1
minus05
0
05
1
15
2
25
(b)
Figure 4 Two-dimensional feature representation for 500 patches of each brain tissue by (a) intensity + PCA and (b) intensity + three-hidden-layer SAE + PCA here PCA is just for visualization of principle components
(a) (b) (c)
Figure 5 Visualization of learned high-level features of input pixel intensities with three-layer SAE (a) (b) and (c) Learned featurerepresentation in the first (with 225 units) second (with 144 units) and third (with 81 units) hidden layers respectively where the features inthe third layer are discriminative for image segmentation task
coding do [31] While SAE can learn higher-level featurescorresponding to the patterns in the appearance of featuresin the former layer (as shown in Figure 5(c)) these high-levelfeatures are more discriminative for image segmentation taskin this paper Hence our model employs the SAE insteadof both hand-craft features and learning based features bysinger-layer AE to extract high-level feature representationfor segmenting brain tissues
233 SAE Plus Softmax for CVH Brain Tissues SegmentationFor every foreground pixel in the cryosection image weextract two patches centered at this pixel from its B-channel(in RGB color space) and V-channel (in HSV color space)image respectively So 120598 in Figure 2 is 2 The two patchesare concatenated together as the input features of SAE andthe features learned by SAE then are sent to a supervised
Softmax classifier The parameters of two SAEs in MODEL1 and MODEL 2 are roughly consistent The patch sizeis set via cross-validation and layer depth is set among1 2 3 considering the balance between computation costand discriminative power The number of units in each layerof SAE is experimentally set as 400 200 and 100 respectivelyThus the final dimensionality of deep-learned feature is 100The weight decay 120582 of two SAEs in MODEL 1 and MODEL 2is set to 0003 and 0005 respectively which is tuned on thevalidation set
234 Ground Truth and Training Sets Generation Theobjec-tive of the proposed model is to automatically segment thethree brain tissues of thewhole 422CVHbrain image at handIt is laborious and time-consuming to manually segmentall the brain image for an anatomical expert so the expert
6 BioMed Research International
Table 1 Details of the SAE architectures with different input patch size and layer depth in this study
only segmented eight CVH brain images which is saturatedwith abundant anatomical structures as the ground truth fortraining sets generation and quantitative evaluation
The patches used for training and testing are extractedfrom the eight labeled images The size of each patch ischosen 17 times 17 to achieve relatively best performance whichis validated in Section 31 so 120596 in Figure 2 is 17 For MODEL1 our training sets contain 10000WM patches and 90000non-WM patches extracted from eight training images Formodel 2 the training sets contain 106000GM patches and84000CSF patches These patches are used for SAE andSoftmax training in the two models
3 Experimental Results and Discussion
In the experiments we firstly focus on evaluating the seg-mentation performance of features learnt by different SAEsbased on cryosection image in order to get the best SAEarchitecture Secondly we compare performances of theproposedmodel based on the deep-learned features and somefamous hand-crafted features such as intensity PCA andHOGand one learning-based feature byAEThen we presenttypical segmentation examples and make some discussionFinally we build the 3D meshes of three tissues based on oursegmentation results
31 Comparison ofDifferent SAEArchitectures Thenonlinearmapping between the input and output of SAE is influencedby its multilayer architecture with various input patch sizesand depths In order to investigate the impact of different SAEarchitectures on segmentation accuracy five different SAEarchitectures are designed and resort to segmentation taskThe detailed parameter configurations are shown in Table 1and the segmentation performances of WM GM and CSFare reported in Table 2
It can be observed from the results that the predictiveaccuracies are generally higher for the architectures withinput patch sizes of 17 times 17 and 21 times 21 The SAEs with
Table 2 Mean and standard deviation of Dice ratio (in ) formeasuring the performance of the three tissue types with fivedifferent architectures trained by using different patch sizes of 5 times 59 times 9 13 times 13 17 times 17 and 21 times 21 respectively The experimentswere conducted in a leave-one-out manner and eight test resultswere collected for each tissue
larger patch size tend to have a deeper hierarchical structureand more trainable parameters These learned parametersare capable of capturing the complex relationship betweeninput and output We can also observe that the architecturewith input patch size of 21 times 21 does not generate substan-tially higher performance suggesting that larger patch mayintroducemore image noise and fused anatomical structuresIn order to obtain better segmentation performance in thefollowing we focus on evaluating the performance of our SAEarchitecture with input patch size of 17 times 17 and depth of 3
32 Comparison of Performances Based on Different Fea-tures In order to provide a comprehensive evaluation ofthe proposed method and illustrate the effect of high-levelfeatures in contrast to low-level features three representativehand-crafted features such as intensity PCA [32 33] andHOG [34] and one learning-based features by AE are usedfor comparison These features follow the same segmentingprocedures as the deep-learned features All the segmentationperformances are reported in Table 3 using Dice ratio It canbe observed that our model with the features extracted bySAE outperforms other well-known features for segmentingall three types of brain tissue Specifically our model yields
BioMed Research International 7
(a)
(b)
(c)
(d)
(e)
Figure 6 Continued
8 BioMed Research International
(f)
(g)
Figure 6 Comparison of the segmentation results with the manually generated segmentation on a cryosection image in CVH dataset (a)shows the original cryosection and its B-channel (in RGB color space) and V-channel (in HSV color space) image (b) shows the manualresults (CSF GM and WM) (c)ndash(g) show the results by the features of our SAE deep learning Intensity PCA HOG and AE respectively
Table 3 Mean and standard deviation of Dice ratio (in ) for thesegmentations obtained by five feature representation methods
the best average Dice ratio of 9069 plusmn 214 (CSF) 9124plusmn 201 (GM) and 9612 plusmn 123 (WM) in a leave-one-outevaluation manner These results have illustrated the strongdiscriminative power of the deep-learned features in braintissues segmentation task
To further demonstrate the advantages of our proposedmodel we visually examine the segmentation results on onecryosection image which is shown in Figure 6 (a) shows theoriginal RGB cryosection image and its B- and V-channelimagesThe ground truth that segmented by experts is shownin (b) (c)ndash(g) present segmentation results of four methodsbased on deep-learned features intensity features PCAfeatures HOG features and AE feautres respectivelyWe cansee that the segmentation results of the proposed model arequite close to the ground truth In contrast other resultseither generate much oversegmentation or fail to segmenttiny anatomical structures accurately Specifically WM isonly adjacent to the GM and can be easily distinguished
from its surroundings so the WM segmentation DRs areapproximate to each other And the visible results of WMare similar in appearance except the fact that HOG-basedmethod mistakenly introduced a small fraction of CSFinto the results It has some difficulty on GM and CSFsegmentation due to the complex anatomical structures andlow contrast HOG and intensity based methods introducemore surroundingnon-GMornon-CSF tissue into theROI ofGM or CSF respectively thus they produce more defects andfuzzy boundaries for different tissues In contrast theGMandCSF tissues generated by ourmethod can be clearly identifiedwith a certain ROI and distinct contours
We then applied our proposed model to segment all422 brain cryosection images Typical images in coronal andsagittal viewpoints and their corresponding segmentationresults (WM GM and CSF) are shown in Figures 7 and 8respectively It is remarkably seen that the results of WMand GM change continuously and their morphological dis-tributions are shown clearly most tiny anatomical structuresare well reserved It is also noticed that the results of CSFseem incomplete and not distributed uniformly This fact isdetermined by the characteristics of the cryosection imagesThese images have such high spatial resolution (0167mmper pixel and 025mm per slice) that it can express finestructures of tissues But since these images were collectedfrom a cadaver the CSF in the brain no longer flowed invivo Due to the effect of gravity the CSF will gather to someplaces rather than uniformly distributing around the surfaceof brain as in live statusThese factors cause the discontinuousdistribution of the CSF segmentation results
BioMed Research International 9
(a)
(b)
(c)
(d)
Figure 7 Coronal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
(a)
(b)
(c)
(d)
Figure 8 Sagittal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
10 BioMed Research International
Figure 9 Three-dimensional surface-rendering reconstruction results of WM GM and CSF based on our segmented images
33 Three-Dimensional Reconstruction Results The CVHcryosection slices are 0167mm per pixel and 025mm thickand such high resolution is very helpful for displaying subtleanatomical structures of brain tissues For a more in-depthunderstanding of these tissues the segmented white mattergray matter and cerebrospinal fluid images are reconstructedusingmarching cubes algorithm to produce 3D surfacemeshThe reconstruction results in different views are shown inFigure 9 From the surface-rendering reconstruction resultsit is seen that the surface of WM is smooth and its sulci andfissures are clearly displayed The distribution of GM shapeis also noticed but the surface of GM reconstruction resultsdoes not look very smooth The reason for this lies in thefact that the GM and CSF are mixed together because of theice crystals in the frozen brain slices so the segmentation ofGM is influenced by its surrounding CSF In spite of it thesulci and cerebral cisterns are also easy to be recognized in3D reconstructed WM and GM
Benefitting from the increasing development of the 3Dreconstruction technology 3D MRI and PET have nowbeen used in clinic and researches But because of theresolution limitation and complexity of brain structures ofthe 2D radiological images (such as CT and MRI) the 3Dreconstruction results are usually unsatisfactory and are hardfor the guidance of clinic operation For the work in ourpaper we focus on the segmentation and reconstruction ofthree kinds of CVHbrain tissuesTheCVHbrain images havehigh spatial resolution of 0167mmper pixel and 025mmperslice after segmentation by the proposed method the high-quality and high-accuracy 2D brain tissues can get well 3Dreconstruction results Some potential applications of the 3Dreconstruction result include the following
(1) The 3D results can be viewed in any orientationbesides the common coronal sagittal and transverseorientations These 3D models (especially for WM
BioMed Research International 11
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
Figure 1 Preprocessing example of cryosection brain image (a) Original imagewithout any preprocessing (3072times 2048 pixels) (b) Croppedimage (1252 times 1364 pixels) (c) Skull stripped image
SAE
Input patches Input patches
Softmax
3-hidden-layer SAE
Softmax
MODEL 1 MODEL 2
SAE
Ground truth Ground truth
3-hidden-layer SAE
(120596 times 120596 times 120598) (120596 times 120596 times 120598)W1 W2
W3
W4
W1 W2
W3 W4
minus
Training MODEL 2 towards GM and CSFsegmentation
Application of segmentation
Training MODEL 1 towards WM segmentation
Figure 2 Flowchart of our segmentation model
The images in CVH dataset are taken of the frozencadaver A total of 422 cross-sectional images of the head(number 1074 to number 1495) are selected for this study Asshown in Figure 1 the slice is 0167mm per pixel 025mmthick and photographed at a resolution of 6291456 (3072times 2048) pixels with 24-bit color information in tiff format[28] In order to reduce computational cost and memoryusage these images are transformed into PNG format andcropped to 1252 times 1364 pixels In the preprocessing stageskull stripping is applied to each image
22 Method Overview In this work the CVH brain tissuesegmentation problem is formulated as a patch classificationtask and the architecture of our segmentationmodel is shownin Figure 2 The model takes patches extracted from the B-channel and V-channel of the original images as input thenSAE is used to extract intrinsic feature representation of the
input patches and the following Softmax classifier generates alabels distribution of these patches based on the deep featuresThis model on segmenting three brain tissues (CSF GMand WM) actually contains two submodels MODEL 1 andMODEL 2 In MODEL 1 GM and CSF are labeled as thesame class and the segmentation is formulated as a three-class classification task WM GM amp CSF and backgroundThrough MODEL 1 WM tissue can be extracted from theregion of interest This segmentation result is helpful to fillthe areas of WM into background so as to eliminate theinfluence of WM So in MODEL 2 the patches from theWM-eliminated image are taken as inputs and the image issegmented into CSF GM and background This pipeline hasthe advantages in that more image patches of the objects withfewer labeled data can be taken as it is quite time-consumingto manually label an image In the following we will describethe details of the model
4 BioMed Research International
W1
W2
W3
EncoderDecoder
h1
n h2
n
h3
n
xn
Figure 3 Proposed three-hidden-layer SAE Note that the number of layers in our model is set via cross-validation
23 Learning Hierarchical Feature Representation by SAE
231 Single-Layer AE A single-layer AE [29] is a kind ofunsupervised neural network whose goal is to minimizethe reconstruction error from inputs to outputs via twocomponents encoder and decoder [19] In the encodingstage given an input sample
119899isin R119873 AE will map it to the
hidden activation ℎ119899isin R119872 by the following mapping
ℎ119899= 119891 (W
1119899+1198871) (1)
where 119891(119911) = 1(1 + exp(minus119911)) is a nonlinear activationfunction W
1isin R119872times119873 is the encoder weight matrix 119887
1isin
R119872 is the bias vector Generally119872 lt 119873 then the networkis forced to learn a compressed representation of the inputvector This compressed representation can be viewed asfeatures of the input vector In the decoding stage AE willreproduce input data from the hidden activation ℎ
119899isin R119872 by
119910119899= 119891 (W
2
ℎ119899+1198872) (2)
whereW2isin R119873times119872 is the decode weight matrix and 119887
2isin R119873
is the bias vectorIn our model during the training stage we minimize the
objective function shown in (3) with respect to the weightmatrixesW
1andW
2and bias vectors 119887
1and 1198872 The objective
function includes an average sum-of-square error term tofit the input data and a weight decay term to decrease themagnitude of weight matrices as well as helping prevent over-fitting
119869 (W b) = 1119898
119898
sum
119899=1
(
1
2
1003817100381710038171003817x119899minus y119899
1003817100381710038171003817
2
)
+
120582
2
119899119897minus1
sum
119897=1
119904119897
sum
119894=1
119904119897+1
sum
119895=1
(119882(119897)
119895119894)
2
(3)
where x119899denotes the 119899-th sample in the training set y
119899
denotes the reconstructed output with input of x119899 120582 denotes
weight decay parameter which controls the relative impor-tance of the two terms 119899
119897denotes the number of layers in the
network 119904119897denotes the number of units in the 119897th layer and
119898 denotes the number of training samples
232 SAE for Hierarchical Feature Learning SAE is a typeof neural network consisting of multiple layers of AEs inwhich the output of each layer is wired to the inputs of thesuccessive layer In this paper we propose a multi-hiddenlayer SAE which is shown in Figure 3 It is noted that thenumber of layers in our model is set via cross-validationFor an input vector
119899 the first layer transforms it into a
vector ℎ1119899that consists of activations of hidden units and the
second layer takes ℎ1119899as input to produce a new activated
vector ℎ2119899 then the final activated vector ℎ3
119899that is produced
by ℎ2119899can be viewed as deep-learned feature representation
of the input sample It is noticed that the model intrinsicallyhandles varying-dimension images through image patcheswith different sizes For a specific task the parameters areusually settled through experiments or experience and thetraining and application of SAE will go on those parameters
In our task we follow the greedy layer-wise trainingstrategy [18 20 30] to obtain better parameters of a SAEThatis we first train the first single AE on the raw input and thentrain the second AE on the hidden activation vector acquiredby the former AE The subsequent layers are repeated usingthe output of each layer as input Once this phase of trainingis complete we stack AEs into SAE and train the entirenetwork by a gradient-based optimization method to refinethe parameters
The high-level features learned by SAE are more discrim-inative compared to hand-crafted feature such as intensityand learning-based feature by single-layer AE To make anintuitive interpretation we conducted a dimension reductionexperiment to visually examine the distributions of featurevectors from image patches by original intensity and a SAEwith three hidden layers respectivelyThe experimental resultis shown in Figure 4 where the dimensionality of each featurevector is reduced to two by Principal Components Analysis(PCA) for the purpose of visualization We can see thatthe features extracted by SAE output a better cluster resultthan intensity features It is easier to generate a separationhyperplane for separating different types of samples
To further visualize the discriminative ability betweenAE and SAE the features learned by each layer of SAEbased on cryosection image are shown in Figure 5 As shownin Figure 5(a) it is seen that AE can only learn primitiveoriented edge-like features just like119870-means ICA or sparse
BioMed Research International 5
WMGMCSF
0 10 20 30minus20 minus10minus30minus15
minus10
minus5
0
5
10
15
(a)
WMGMCSF
0 1 2 3 4minus2minus3minus4 minus1minus5minus2
minus15
minus1
minus05
0
05
1
15
2
25
(b)
Figure 4 Two-dimensional feature representation for 500 patches of each brain tissue by (a) intensity + PCA and (b) intensity + three-hidden-layer SAE + PCA here PCA is just for visualization of principle components
(a) (b) (c)
Figure 5 Visualization of learned high-level features of input pixel intensities with three-layer SAE (a) (b) and (c) Learned featurerepresentation in the first (with 225 units) second (with 144 units) and third (with 81 units) hidden layers respectively where the features inthe third layer are discriminative for image segmentation task
coding do [31] While SAE can learn higher-level featurescorresponding to the patterns in the appearance of featuresin the former layer (as shown in Figure 5(c)) these high-levelfeatures are more discriminative for image segmentation taskin this paper Hence our model employs the SAE insteadof both hand-craft features and learning based features bysinger-layer AE to extract high-level feature representationfor segmenting brain tissues
233 SAE Plus Softmax for CVH Brain Tissues SegmentationFor every foreground pixel in the cryosection image weextract two patches centered at this pixel from its B-channel(in RGB color space) and V-channel (in HSV color space)image respectively So 120598 in Figure 2 is 2 The two patchesare concatenated together as the input features of SAE andthe features learned by SAE then are sent to a supervised
Softmax classifier The parameters of two SAEs in MODEL1 and MODEL 2 are roughly consistent The patch sizeis set via cross-validation and layer depth is set among1 2 3 considering the balance between computation costand discriminative power The number of units in each layerof SAE is experimentally set as 400 200 and 100 respectivelyThus the final dimensionality of deep-learned feature is 100The weight decay 120582 of two SAEs in MODEL 1 and MODEL 2is set to 0003 and 0005 respectively which is tuned on thevalidation set
234 Ground Truth and Training Sets Generation Theobjec-tive of the proposed model is to automatically segment thethree brain tissues of thewhole 422CVHbrain image at handIt is laborious and time-consuming to manually segmentall the brain image for an anatomical expert so the expert
6 BioMed Research International
Table 1 Details of the SAE architectures with different input patch size and layer depth in this study
only segmented eight CVH brain images which is saturatedwith abundant anatomical structures as the ground truth fortraining sets generation and quantitative evaluation
The patches used for training and testing are extractedfrom the eight labeled images The size of each patch ischosen 17 times 17 to achieve relatively best performance whichis validated in Section 31 so 120596 in Figure 2 is 17 For MODEL1 our training sets contain 10000WM patches and 90000non-WM patches extracted from eight training images Formodel 2 the training sets contain 106000GM patches and84000CSF patches These patches are used for SAE andSoftmax training in the two models
3 Experimental Results and Discussion
In the experiments we firstly focus on evaluating the seg-mentation performance of features learnt by different SAEsbased on cryosection image in order to get the best SAEarchitecture Secondly we compare performances of theproposedmodel based on the deep-learned features and somefamous hand-crafted features such as intensity PCA andHOGand one learning-based feature byAEThen we presenttypical segmentation examples and make some discussionFinally we build the 3D meshes of three tissues based on oursegmentation results
31 Comparison ofDifferent SAEArchitectures Thenonlinearmapping between the input and output of SAE is influencedby its multilayer architecture with various input patch sizesand depths In order to investigate the impact of different SAEarchitectures on segmentation accuracy five different SAEarchitectures are designed and resort to segmentation taskThe detailed parameter configurations are shown in Table 1and the segmentation performances of WM GM and CSFare reported in Table 2
It can be observed from the results that the predictiveaccuracies are generally higher for the architectures withinput patch sizes of 17 times 17 and 21 times 21 The SAEs with
Table 2 Mean and standard deviation of Dice ratio (in ) formeasuring the performance of the three tissue types with fivedifferent architectures trained by using different patch sizes of 5 times 59 times 9 13 times 13 17 times 17 and 21 times 21 respectively The experimentswere conducted in a leave-one-out manner and eight test resultswere collected for each tissue
larger patch size tend to have a deeper hierarchical structureand more trainable parameters These learned parametersare capable of capturing the complex relationship betweeninput and output We can also observe that the architecturewith input patch size of 21 times 21 does not generate substan-tially higher performance suggesting that larger patch mayintroducemore image noise and fused anatomical structuresIn order to obtain better segmentation performance in thefollowing we focus on evaluating the performance of our SAEarchitecture with input patch size of 17 times 17 and depth of 3
32 Comparison of Performances Based on Different Fea-tures In order to provide a comprehensive evaluation ofthe proposed method and illustrate the effect of high-levelfeatures in contrast to low-level features three representativehand-crafted features such as intensity PCA [32 33] andHOG [34] and one learning-based features by AE are usedfor comparison These features follow the same segmentingprocedures as the deep-learned features All the segmentationperformances are reported in Table 3 using Dice ratio It canbe observed that our model with the features extracted bySAE outperforms other well-known features for segmentingall three types of brain tissue Specifically our model yields
BioMed Research International 7
(a)
(b)
(c)
(d)
(e)
Figure 6 Continued
8 BioMed Research International
(f)
(g)
Figure 6 Comparison of the segmentation results with the manually generated segmentation on a cryosection image in CVH dataset (a)shows the original cryosection and its B-channel (in RGB color space) and V-channel (in HSV color space) image (b) shows the manualresults (CSF GM and WM) (c)ndash(g) show the results by the features of our SAE deep learning Intensity PCA HOG and AE respectively
Table 3 Mean and standard deviation of Dice ratio (in ) for thesegmentations obtained by five feature representation methods
the best average Dice ratio of 9069 plusmn 214 (CSF) 9124plusmn 201 (GM) and 9612 plusmn 123 (WM) in a leave-one-outevaluation manner These results have illustrated the strongdiscriminative power of the deep-learned features in braintissues segmentation task
To further demonstrate the advantages of our proposedmodel we visually examine the segmentation results on onecryosection image which is shown in Figure 6 (a) shows theoriginal RGB cryosection image and its B- and V-channelimagesThe ground truth that segmented by experts is shownin (b) (c)ndash(g) present segmentation results of four methodsbased on deep-learned features intensity features PCAfeatures HOG features and AE feautres respectivelyWe cansee that the segmentation results of the proposed model arequite close to the ground truth In contrast other resultseither generate much oversegmentation or fail to segmenttiny anatomical structures accurately Specifically WM isonly adjacent to the GM and can be easily distinguished
from its surroundings so the WM segmentation DRs areapproximate to each other And the visible results of WMare similar in appearance except the fact that HOG-basedmethod mistakenly introduced a small fraction of CSFinto the results It has some difficulty on GM and CSFsegmentation due to the complex anatomical structures andlow contrast HOG and intensity based methods introducemore surroundingnon-GMornon-CSF tissue into theROI ofGM or CSF respectively thus they produce more defects andfuzzy boundaries for different tissues In contrast theGMandCSF tissues generated by ourmethod can be clearly identifiedwith a certain ROI and distinct contours
We then applied our proposed model to segment all422 brain cryosection images Typical images in coronal andsagittal viewpoints and their corresponding segmentationresults (WM GM and CSF) are shown in Figures 7 and 8respectively It is remarkably seen that the results of WMand GM change continuously and their morphological dis-tributions are shown clearly most tiny anatomical structuresare well reserved It is also noticed that the results of CSFseem incomplete and not distributed uniformly This fact isdetermined by the characteristics of the cryosection imagesThese images have such high spatial resolution (0167mmper pixel and 025mm per slice) that it can express finestructures of tissues But since these images were collectedfrom a cadaver the CSF in the brain no longer flowed invivo Due to the effect of gravity the CSF will gather to someplaces rather than uniformly distributing around the surfaceof brain as in live statusThese factors cause the discontinuousdistribution of the CSF segmentation results
BioMed Research International 9
(a)
(b)
(c)
(d)
Figure 7 Coronal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
(a)
(b)
(c)
(d)
Figure 8 Sagittal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
10 BioMed Research International
Figure 9 Three-dimensional surface-rendering reconstruction results of WM GM and CSF based on our segmented images
33 Three-Dimensional Reconstruction Results The CVHcryosection slices are 0167mm per pixel and 025mm thickand such high resolution is very helpful for displaying subtleanatomical structures of brain tissues For a more in-depthunderstanding of these tissues the segmented white mattergray matter and cerebrospinal fluid images are reconstructedusingmarching cubes algorithm to produce 3D surfacemeshThe reconstruction results in different views are shown inFigure 9 From the surface-rendering reconstruction resultsit is seen that the surface of WM is smooth and its sulci andfissures are clearly displayed The distribution of GM shapeis also noticed but the surface of GM reconstruction resultsdoes not look very smooth The reason for this lies in thefact that the GM and CSF are mixed together because of theice crystals in the frozen brain slices so the segmentation ofGM is influenced by its surrounding CSF In spite of it thesulci and cerebral cisterns are also easy to be recognized in3D reconstructed WM and GM
Benefitting from the increasing development of the 3Dreconstruction technology 3D MRI and PET have nowbeen used in clinic and researches But because of theresolution limitation and complexity of brain structures ofthe 2D radiological images (such as CT and MRI) the 3Dreconstruction results are usually unsatisfactory and are hardfor the guidance of clinic operation For the work in ourpaper we focus on the segmentation and reconstruction ofthree kinds of CVHbrain tissuesTheCVHbrain images havehigh spatial resolution of 0167mmper pixel and 025mmperslice after segmentation by the proposed method the high-quality and high-accuracy 2D brain tissues can get well 3Dreconstruction results Some potential applications of the 3Dreconstruction result include the following
(1) The 3D results can be viewed in any orientationbesides the common coronal sagittal and transverseorientations These 3D models (especially for WM
BioMed Research International 11
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
Figure 3 Proposed three-hidden-layer SAE Note that the number of layers in our model is set via cross-validation
23 Learning Hierarchical Feature Representation by SAE
231 Single-Layer AE A single-layer AE [29] is a kind ofunsupervised neural network whose goal is to minimizethe reconstruction error from inputs to outputs via twocomponents encoder and decoder [19] In the encodingstage given an input sample
119899isin R119873 AE will map it to the
hidden activation ℎ119899isin R119872 by the following mapping
ℎ119899= 119891 (W
1119899+1198871) (1)
where 119891(119911) = 1(1 + exp(minus119911)) is a nonlinear activationfunction W
1isin R119872times119873 is the encoder weight matrix 119887
1isin
R119872 is the bias vector Generally119872 lt 119873 then the networkis forced to learn a compressed representation of the inputvector This compressed representation can be viewed asfeatures of the input vector In the decoding stage AE willreproduce input data from the hidden activation ℎ
119899isin R119872 by
119910119899= 119891 (W
2
ℎ119899+1198872) (2)
whereW2isin R119873times119872 is the decode weight matrix and 119887
2isin R119873
is the bias vectorIn our model during the training stage we minimize the
objective function shown in (3) with respect to the weightmatrixesW
1andW
2and bias vectors 119887
1and 1198872 The objective
function includes an average sum-of-square error term tofit the input data and a weight decay term to decrease themagnitude of weight matrices as well as helping prevent over-fitting
119869 (W b) = 1119898
119898
sum
119899=1
(
1
2
1003817100381710038171003817x119899minus y119899
1003817100381710038171003817
2
)
+
120582
2
119899119897minus1
sum
119897=1
119904119897
sum
119894=1
119904119897+1
sum
119895=1
(119882(119897)
119895119894)
2
(3)
where x119899denotes the 119899-th sample in the training set y
119899
denotes the reconstructed output with input of x119899 120582 denotes
weight decay parameter which controls the relative impor-tance of the two terms 119899
119897denotes the number of layers in the
network 119904119897denotes the number of units in the 119897th layer and
119898 denotes the number of training samples
232 SAE for Hierarchical Feature Learning SAE is a typeof neural network consisting of multiple layers of AEs inwhich the output of each layer is wired to the inputs of thesuccessive layer In this paper we propose a multi-hiddenlayer SAE which is shown in Figure 3 It is noted that thenumber of layers in our model is set via cross-validationFor an input vector
119899 the first layer transforms it into a
vector ℎ1119899that consists of activations of hidden units and the
second layer takes ℎ1119899as input to produce a new activated
vector ℎ2119899 then the final activated vector ℎ3
119899that is produced
by ℎ2119899can be viewed as deep-learned feature representation
of the input sample It is noticed that the model intrinsicallyhandles varying-dimension images through image patcheswith different sizes For a specific task the parameters areusually settled through experiments or experience and thetraining and application of SAE will go on those parameters
In our task we follow the greedy layer-wise trainingstrategy [18 20 30] to obtain better parameters of a SAEThatis we first train the first single AE on the raw input and thentrain the second AE on the hidden activation vector acquiredby the former AE The subsequent layers are repeated usingthe output of each layer as input Once this phase of trainingis complete we stack AEs into SAE and train the entirenetwork by a gradient-based optimization method to refinethe parameters
The high-level features learned by SAE are more discrim-inative compared to hand-crafted feature such as intensityand learning-based feature by single-layer AE To make anintuitive interpretation we conducted a dimension reductionexperiment to visually examine the distributions of featurevectors from image patches by original intensity and a SAEwith three hidden layers respectivelyThe experimental resultis shown in Figure 4 where the dimensionality of each featurevector is reduced to two by Principal Components Analysis(PCA) for the purpose of visualization We can see thatthe features extracted by SAE output a better cluster resultthan intensity features It is easier to generate a separationhyperplane for separating different types of samples
To further visualize the discriminative ability betweenAE and SAE the features learned by each layer of SAEbased on cryosection image are shown in Figure 5 As shownin Figure 5(a) it is seen that AE can only learn primitiveoriented edge-like features just like119870-means ICA or sparse
BioMed Research International 5
WMGMCSF
0 10 20 30minus20 minus10minus30minus15
minus10
minus5
0
5
10
15
(a)
WMGMCSF
0 1 2 3 4minus2minus3minus4 minus1minus5minus2
minus15
minus1
minus05
0
05
1
15
2
25
(b)
Figure 4 Two-dimensional feature representation for 500 patches of each brain tissue by (a) intensity + PCA and (b) intensity + three-hidden-layer SAE + PCA here PCA is just for visualization of principle components
(a) (b) (c)
Figure 5 Visualization of learned high-level features of input pixel intensities with three-layer SAE (a) (b) and (c) Learned featurerepresentation in the first (with 225 units) second (with 144 units) and third (with 81 units) hidden layers respectively where the features inthe third layer are discriminative for image segmentation task
coding do [31] While SAE can learn higher-level featurescorresponding to the patterns in the appearance of featuresin the former layer (as shown in Figure 5(c)) these high-levelfeatures are more discriminative for image segmentation taskin this paper Hence our model employs the SAE insteadof both hand-craft features and learning based features bysinger-layer AE to extract high-level feature representationfor segmenting brain tissues
233 SAE Plus Softmax for CVH Brain Tissues SegmentationFor every foreground pixel in the cryosection image weextract two patches centered at this pixel from its B-channel(in RGB color space) and V-channel (in HSV color space)image respectively So 120598 in Figure 2 is 2 The two patchesare concatenated together as the input features of SAE andthe features learned by SAE then are sent to a supervised
Softmax classifier The parameters of two SAEs in MODEL1 and MODEL 2 are roughly consistent The patch sizeis set via cross-validation and layer depth is set among1 2 3 considering the balance between computation costand discriminative power The number of units in each layerof SAE is experimentally set as 400 200 and 100 respectivelyThus the final dimensionality of deep-learned feature is 100The weight decay 120582 of two SAEs in MODEL 1 and MODEL 2is set to 0003 and 0005 respectively which is tuned on thevalidation set
234 Ground Truth and Training Sets Generation Theobjec-tive of the proposed model is to automatically segment thethree brain tissues of thewhole 422CVHbrain image at handIt is laborious and time-consuming to manually segmentall the brain image for an anatomical expert so the expert
6 BioMed Research International
Table 1 Details of the SAE architectures with different input patch size and layer depth in this study
only segmented eight CVH brain images which is saturatedwith abundant anatomical structures as the ground truth fortraining sets generation and quantitative evaluation
The patches used for training and testing are extractedfrom the eight labeled images The size of each patch ischosen 17 times 17 to achieve relatively best performance whichis validated in Section 31 so 120596 in Figure 2 is 17 For MODEL1 our training sets contain 10000WM patches and 90000non-WM patches extracted from eight training images Formodel 2 the training sets contain 106000GM patches and84000CSF patches These patches are used for SAE andSoftmax training in the two models
3 Experimental Results and Discussion
In the experiments we firstly focus on evaluating the seg-mentation performance of features learnt by different SAEsbased on cryosection image in order to get the best SAEarchitecture Secondly we compare performances of theproposedmodel based on the deep-learned features and somefamous hand-crafted features such as intensity PCA andHOGand one learning-based feature byAEThen we presenttypical segmentation examples and make some discussionFinally we build the 3D meshes of three tissues based on oursegmentation results
31 Comparison ofDifferent SAEArchitectures Thenonlinearmapping between the input and output of SAE is influencedby its multilayer architecture with various input patch sizesand depths In order to investigate the impact of different SAEarchitectures on segmentation accuracy five different SAEarchitectures are designed and resort to segmentation taskThe detailed parameter configurations are shown in Table 1and the segmentation performances of WM GM and CSFare reported in Table 2
It can be observed from the results that the predictiveaccuracies are generally higher for the architectures withinput patch sizes of 17 times 17 and 21 times 21 The SAEs with
Table 2 Mean and standard deviation of Dice ratio (in ) formeasuring the performance of the three tissue types with fivedifferent architectures trained by using different patch sizes of 5 times 59 times 9 13 times 13 17 times 17 and 21 times 21 respectively The experimentswere conducted in a leave-one-out manner and eight test resultswere collected for each tissue
larger patch size tend to have a deeper hierarchical structureand more trainable parameters These learned parametersare capable of capturing the complex relationship betweeninput and output We can also observe that the architecturewith input patch size of 21 times 21 does not generate substan-tially higher performance suggesting that larger patch mayintroducemore image noise and fused anatomical structuresIn order to obtain better segmentation performance in thefollowing we focus on evaluating the performance of our SAEarchitecture with input patch size of 17 times 17 and depth of 3
32 Comparison of Performances Based on Different Fea-tures In order to provide a comprehensive evaluation ofthe proposed method and illustrate the effect of high-levelfeatures in contrast to low-level features three representativehand-crafted features such as intensity PCA [32 33] andHOG [34] and one learning-based features by AE are usedfor comparison These features follow the same segmentingprocedures as the deep-learned features All the segmentationperformances are reported in Table 3 using Dice ratio It canbe observed that our model with the features extracted bySAE outperforms other well-known features for segmentingall three types of brain tissue Specifically our model yields
BioMed Research International 7
(a)
(b)
(c)
(d)
(e)
Figure 6 Continued
8 BioMed Research International
(f)
(g)
Figure 6 Comparison of the segmentation results with the manually generated segmentation on a cryosection image in CVH dataset (a)shows the original cryosection and its B-channel (in RGB color space) and V-channel (in HSV color space) image (b) shows the manualresults (CSF GM and WM) (c)ndash(g) show the results by the features of our SAE deep learning Intensity PCA HOG and AE respectively
Table 3 Mean and standard deviation of Dice ratio (in ) for thesegmentations obtained by five feature representation methods
the best average Dice ratio of 9069 plusmn 214 (CSF) 9124plusmn 201 (GM) and 9612 plusmn 123 (WM) in a leave-one-outevaluation manner These results have illustrated the strongdiscriminative power of the deep-learned features in braintissues segmentation task
To further demonstrate the advantages of our proposedmodel we visually examine the segmentation results on onecryosection image which is shown in Figure 6 (a) shows theoriginal RGB cryosection image and its B- and V-channelimagesThe ground truth that segmented by experts is shownin (b) (c)ndash(g) present segmentation results of four methodsbased on deep-learned features intensity features PCAfeatures HOG features and AE feautres respectivelyWe cansee that the segmentation results of the proposed model arequite close to the ground truth In contrast other resultseither generate much oversegmentation or fail to segmenttiny anatomical structures accurately Specifically WM isonly adjacent to the GM and can be easily distinguished
from its surroundings so the WM segmentation DRs areapproximate to each other And the visible results of WMare similar in appearance except the fact that HOG-basedmethod mistakenly introduced a small fraction of CSFinto the results It has some difficulty on GM and CSFsegmentation due to the complex anatomical structures andlow contrast HOG and intensity based methods introducemore surroundingnon-GMornon-CSF tissue into theROI ofGM or CSF respectively thus they produce more defects andfuzzy boundaries for different tissues In contrast theGMandCSF tissues generated by ourmethod can be clearly identifiedwith a certain ROI and distinct contours
We then applied our proposed model to segment all422 brain cryosection images Typical images in coronal andsagittal viewpoints and their corresponding segmentationresults (WM GM and CSF) are shown in Figures 7 and 8respectively It is remarkably seen that the results of WMand GM change continuously and their morphological dis-tributions are shown clearly most tiny anatomical structuresare well reserved It is also noticed that the results of CSFseem incomplete and not distributed uniformly This fact isdetermined by the characteristics of the cryosection imagesThese images have such high spatial resolution (0167mmper pixel and 025mm per slice) that it can express finestructures of tissues But since these images were collectedfrom a cadaver the CSF in the brain no longer flowed invivo Due to the effect of gravity the CSF will gather to someplaces rather than uniformly distributing around the surfaceof brain as in live statusThese factors cause the discontinuousdistribution of the CSF segmentation results
BioMed Research International 9
(a)
(b)
(c)
(d)
Figure 7 Coronal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
(a)
(b)
(c)
(d)
Figure 8 Sagittal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
10 BioMed Research International
Figure 9 Three-dimensional surface-rendering reconstruction results of WM GM and CSF based on our segmented images
33 Three-Dimensional Reconstruction Results The CVHcryosection slices are 0167mm per pixel and 025mm thickand such high resolution is very helpful for displaying subtleanatomical structures of brain tissues For a more in-depthunderstanding of these tissues the segmented white mattergray matter and cerebrospinal fluid images are reconstructedusingmarching cubes algorithm to produce 3D surfacemeshThe reconstruction results in different views are shown inFigure 9 From the surface-rendering reconstruction resultsit is seen that the surface of WM is smooth and its sulci andfissures are clearly displayed The distribution of GM shapeis also noticed but the surface of GM reconstruction resultsdoes not look very smooth The reason for this lies in thefact that the GM and CSF are mixed together because of theice crystals in the frozen brain slices so the segmentation ofGM is influenced by its surrounding CSF In spite of it thesulci and cerebral cisterns are also easy to be recognized in3D reconstructed WM and GM
Benefitting from the increasing development of the 3Dreconstruction technology 3D MRI and PET have nowbeen used in clinic and researches But because of theresolution limitation and complexity of brain structures ofthe 2D radiological images (such as CT and MRI) the 3Dreconstruction results are usually unsatisfactory and are hardfor the guidance of clinic operation For the work in ourpaper we focus on the segmentation and reconstruction ofthree kinds of CVHbrain tissuesTheCVHbrain images havehigh spatial resolution of 0167mmper pixel and 025mmperslice after segmentation by the proposed method the high-quality and high-accuracy 2D brain tissues can get well 3Dreconstruction results Some potential applications of the 3Dreconstruction result include the following
(1) The 3D results can be viewed in any orientationbesides the common coronal sagittal and transverseorientations These 3D models (especially for WM
BioMed Research International 11
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
Figure 4 Two-dimensional feature representation for 500 patches of each brain tissue by (a) intensity + PCA and (b) intensity + three-hidden-layer SAE + PCA here PCA is just for visualization of principle components
(a) (b) (c)
Figure 5 Visualization of learned high-level features of input pixel intensities with three-layer SAE (a) (b) and (c) Learned featurerepresentation in the first (with 225 units) second (with 144 units) and third (with 81 units) hidden layers respectively where the features inthe third layer are discriminative for image segmentation task
coding do [31] While SAE can learn higher-level featurescorresponding to the patterns in the appearance of featuresin the former layer (as shown in Figure 5(c)) these high-levelfeatures are more discriminative for image segmentation taskin this paper Hence our model employs the SAE insteadof both hand-craft features and learning based features bysinger-layer AE to extract high-level feature representationfor segmenting brain tissues
233 SAE Plus Softmax for CVH Brain Tissues SegmentationFor every foreground pixel in the cryosection image weextract two patches centered at this pixel from its B-channel(in RGB color space) and V-channel (in HSV color space)image respectively So 120598 in Figure 2 is 2 The two patchesare concatenated together as the input features of SAE andthe features learned by SAE then are sent to a supervised
Softmax classifier The parameters of two SAEs in MODEL1 and MODEL 2 are roughly consistent The patch sizeis set via cross-validation and layer depth is set among1 2 3 considering the balance between computation costand discriminative power The number of units in each layerof SAE is experimentally set as 400 200 and 100 respectivelyThus the final dimensionality of deep-learned feature is 100The weight decay 120582 of two SAEs in MODEL 1 and MODEL 2is set to 0003 and 0005 respectively which is tuned on thevalidation set
234 Ground Truth and Training Sets Generation Theobjec-tive of the proposed model is to automatically segment thethree brain tissues of thewhole 422CVHbrain image at handIt is laborious and time-consuming to manually segmentall the brain image for an anatomical expert so the expert
6 BioMed Research International
Table 1 Details of the SAE architectures with different input patch size and layer depth in this study
only segmented eight CVH brain images which is saturatedwith abundant anatomical structures as the ground truth fortraining sets generation and quantitative evaluation
The patches used for training and testing are extractedfrom the eight labeled images The size of each patch ischosen 17 times 17 to achieve relatively best performance whichis validated in Section 31 so 120596 in Figure 2 is 17 For MODEL1 our training sets contain 10000WM patches and 90000non-WM patches extracted from eight training images Formodel 2 the training sets contain 106000GM patches and84000CSF patches These patches are used for SAE andSoftmax training in the two models
3 Experimental Results and Discussion
In the experiments we firstly focus on evaluating the seg-mentation performance of features learnt by different SAEsbased on cryosection image in order to get the best SAEarchitecture Secondly we compare performances of theproposedmodel based on the deep-learned features and somefamous hand-crafted features such as intensity PCA andHOGand one learning-based feature byAEThen we presenttypical segmentation examples and make some discussionFinally we build the 3D meshes of three tissues based on oursegmentation results
31 Comparison ofDifferent SAEArchitectures Thenonlinearmapping between the input and output of SAE is influencedby its multilayer architecture with various input patch sizesand depths In order to investigate the impact of different SAEarchitectures on segmentation accuracy five different SAEarchitectures are designed and resort to segmentation taskThe detailed parameter configurations are shown in Table 1and the segmentation performances of WM GM and CSFare reported in Table 2
It can be observed from the results that the predictiveaccuracies are generally higher for the architectures withinput patch sizes of 17 times 17 and 21 times 21 The SAEs with
Table 2 Mean and standard deviation of Dice ratio (in ) formeasuring the performance of the three tissue types with fivedifferent architectures trained by using different patch sizes of 5 times 59 times 9 13 times 13 17 times 17 and 21 times 21 respectively The experimentswere conducted in a leave-one-out manner and eight test resultswere collected for each tissue
larger patch size tend to have a deeper hierarchical structureand more trainable parameters These learned parametersare capable of capturing the complex relationship betweeninput and output We can also observe that the architecturewith input patch size of 21 times 21 does not generate substan-tially higher performance suggesting that larger patch mayintroducemore image noise and fused anatomical structuresIn order to obtain better segmentation performance in thefollowing we focus on evaluating the performance of our SAEarchitecture with input patch size of 17 times 17 and depth of 3
32 Comparison of Performances Based on Different Fea-tures In order to provide a comprehensive evaluation ofthe proposed method and illustrate the effect of high-levelfeatures in contrast to low-level features three representativehand-crafted features such as intensity PCA [32 33] andHOG [34] and one learning-based features by AE are usedfor comparison These features follow the same segmentingprocedures as the deep-learned features All the segmentationperformances are reported in Table 3 using Dice ratio It canbe observed that our model with the features extracted bySAE outperforms other well-known features for segmentingall three types of brain tissue Specifically our model yields
BioMed Research International 7
(a)
(b)
(c)
(d)
(e)
Figure 6 Continued
8 BioMed Research International
(f)
(g)
Figure 6 Comparison of the segmentation results with the manually generated segmentation on a cryosection image in CVH dataset (a)shows the original cryosection and its B-channel (in RGB color space) and V-channel (in HSV color space) image (b) shows the manualresults (CSF GM and WM) (c)ndash(g) show the results by the features of our SAE deep learning Intensity PCA HOG and AE respectively
Table 3 Mean and standard deviation of Dice ratio (in ) for thesegmentations obtained by five feature representation methods
the best average Dice ratio of 9069 plusmn 214 (CSF) 9124plusmn 201 (GM) and 9612 plusmn 123 (WM) in a leave-one-outevaluation manner These results have illustrated the strongdiscriminative power of the deep-learned features in braintissues segmentation task
To further demonstrate the advantages of our proposedmodel we visually examine the segmentation results on onecryosection image which is shown in Figure 6 (a) shows theoriginal RGB cryosection image and its B- and V-channelimagesThe ground truth that segmented by experts is shownin (b) (c)ndash(g) present segmentation results of four methodsbased on deep-learned features intensity features PCAfeatures HOG features and AE feautres respectivelyWe cansee that the segmentation results of the proposed model arequite close to the ground truth In contrast other resultseither generate much oversegmentation or fail to segmenttiny anatomical structures accurately Specifically WM isonly adjacent to the GM and can be easily distinguished
from its surroundings so the WM segmentation DRs areapproximate to each other And the visible results of WMare similar in appearance except the fact that HOG-basedmethod mistakenly introduced a small fraction of CSFinto the results It has some difficulty on GM and CSFsegmentation due to the complex anatomical structures andlow contrast HOG and intensity based methods introducemore surroundingnon-GMornon-CSF tissue into theROI ofGM or CSF respectively thus they produce more defects andfuzzy boundaries for different tissues In contrast theGMandCSF tissues generated by ourmethod can be clearly identifiedwith a certain ROI and distinct contours
We then applied our proposed model to segment all422 brain cryosection images Typical images in coronal andsagittal viewpoints and their corresponding segmentationresults (WM GM and CSF) are shown in Figures 7 and 8respectively It is remarkably seen that the results of WMand GM change continuously and their morphological dis-tributions are shown clearly most tiny anatomical structuresare well reserved It is also noticed that the results of CSFseem incomplete and not distributed uniformly This fact isdetermined by the characteristics of the cryosection imagesThese images have such high spatial resolution (0167mmper pixel and 025mm per slice) that it can express finestructures of tissues But since these images were collectedfrom a cadaver the CSF in the brain no longer flowed invivo Due to the effect of gravity the CSF will gather to someplaces rather than uniformly distributing around the surfaceof brain as in live statusThese factors cause the discontinuousdistribution of the CSF segmentation results
BioMed Research International 9
(a)
(b)
(c)
(d)
Figure 7 Coronal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
(a)
(b)
(c)
(d)
Figure 8 Sagittal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
10 BioMed Research International
Figure 9 Three-dimensional surface-rendering reconstruction results of WM GM and CSF based on our segmented images
33 Three-Dimensional Reconstruction Results The CVHcryosection slices are 0167mm per pixel and 025mm thickand such high resolution is very helpful for displaying subtleanatomical structures of brain tissues For a more in-depthunderstanding of these tissues the segmented white mattergray matter and cerebrospinal fluid images are reconstructedusingmarching cubes algorithm to produce 3D surfacemeshThe reconstruction results in different views are shown inFigure 9 From the surface-rendering reconstruction resultsit is seen that the surface of WM is smooth and its sulci andfissures are clearly displayed The distribution of GM shapeis also noticed but the surface of GM reconstruction resultsdoes not look very smooth The reason for this lies in thefact that the GM and CSF are mixed together because of theice crystals in the frozen brain slices so the segmentation ofGM is influenced by its surrounding CSF In spite of it thesulci and cerebral cisterns are also easy to be recognized in3D reconstructed WM and GM
Benefitting from the increasing development of the 3Dreconstruction technology 3D MRI and PET have nowbeen used in clinic and researches But because of theresolution limitation and complexity of brain structures ofthe 2D radiological images (such as CT and MRI) the 3Dreconstruction results are usually unsatisfactory and are hardfor the guidance of clinic operation For the work in ourpaper we focus on the segmentation and reconstruction ofthree kinds of CVHbrain tissuesTheCVHbrain images havehigh spatial resolution of 0167mmper pixel and 025mmperslice after segmentation by the proposed method the high-quality and high-accuracy 2D brain tissues can get well 3Dreconstruction results Some potential applications of the 3Dreconstruction result include the following
(1) The 3D results can be viewed in any orientationbesides the common coronal sagittal and transverseorientations These 3D models (especially for WM
BioMed Research International 11
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
only segmented eight CVH brain images which is saturatedwith abundant anatomical structures as the ground truth fortraining sets generation and quantitative evaluation
The patches used for training and testing are extractedfrom the eight labeled images The size of each patch ischosen 17 times 17 to achieve relatively best performance whichis validated in Section 31 so 120596 in Figure 2 is 17 For MODEL1 our training sets contain 10000WM patches and 90000non-WM patches extracted from eight training images Formodel 2 the training sets contain 106000GM patches and84000CSF patches These patches are used for SAE andSoftmax training in the two models
3 Experimental Results and Discussion
In the experiments we firstly focus on evaluating the seg-mentation performance of features learnt by different SAEsbased on cryosection image in order to get the best SAEarchitecture Secondly we compare performances of theproposedmodel based on the deep-learned features and somefamous hand-crafted features such as intensity PCA andHOGand one learning-based feature byAEThen we presenttypical segmentation examples and make some discussionFinally we build the 3D meshes of three tissues based on oursegmentation results
31 Comparison ofDifferent SAEArchitectures Thenonlinearmapping between the input and output of SAE is influencedby its multilayer architecture with various input patch sizesand depths In order to investigate the impact of different SAEarchitectures on segmentation accuracy five different SAEarchitectures are designed and resort to segmentation taskThe detailed parameter configurations are shown in Table 1and the segmentation performances of WM GM and CSFare reported in Table 2
It can be observed from the results that the predictiveaccuracies are generally higher for the architectures withinput patch sizes of 17 times 17 and 21 times 21 The SAEs with
Table 2 Mean and standard deviation of Dice ratio (in ) formeasuring the performance of the three tissue types with fivedifferent architectures trained by using different patch sizes of 5 times 59 times 9 13 times 13 17 times 17 and 21 times 21 respectively The experimentswere conducted in a leave-one-out manner and eight test resultswere collected for each tissue
larger patch size tend to have a deeper hierarchical structureand more trainable parameters These learned parametersare capable of capturing the complex relationship betweeninput and output We can also observe that the architecturewith input patch size of 21 times 21 does not generate substan-tially higher performance suggesting that larger patch mayintroducemore image noise and fused anatomical structuresIn order to obtain better segmentation performance in thefollowing we focus on evaluating the performance of our SAEarchitecture with input patch size of 17 times 17 and depth of 3
32 Comparison of Performances Based on Different Fea-tures In order to provide a comprehensive evaluation ofthe proposed method and illustrate the effect of high-levelfeatures in contrast to low-level features three representativehand-crafted features such as intensity PCA [32 33] andHOG [34] and one learning-based features by AE are usedfor comparison These features follow the same segmentingprocedures as the deep-learned features All the segmentationperformances are reported in Table 3 using Dice ratio It canbe observed that our model with the features extracted bySAE outperforms other well-known features for segmentingall three types of brain tissue Specifically our model yields
BioMed Research International 7
(a)
(b)
(c)
(d)
(e)
Figure 6 Continued
8 BioMed Research International
(f)
(g)
Figure 6 Comparison of the segmentation results with the manually generated segmentation on a cryosection image in CVH dataset (a)shows the original cryosection and its B-channel (in RGB color space) and V-channel (in HSV color space) image (b) shows the manualresults (CSF GM and WM) (c)ndash(g) show the results by the features of our SAE deep learning Intensity PCA HOG and AE respectively
Table 3 Mean and standard deviation of Dice ratio (in ) for thesegmentations obtained by five feature representation methods
the best average Dice ratio of 9069 plusmn 214 (CSF) 9124plusmn 201 (GM) and 9612 plusmn 123 (WM) in a leave-one-outevaluation manner These results have illustrated the strongdiscriminative power of the deep-learned features in braintissues segmentation task
To further demonstrate the advantages of our proposedmodel we visually examine the segmentation results on onecryosection image which is shown in Figure 6 (a) shows theoriginal RGB cryosection image and its B- and V-channelimagesThe ground truth that segmented by experts is shownin (b) (c)ndash(g) present segmentation results of four methodsbased on deep-learned features intensity features PCAfeatures HOG features and AE feautres respectivelyWe cansee that the segmentation results of the proposed model arequite close to the ground truth In contrast other resultseither generate much oversegmentation or fail to segmenttiny anatomical structures accurately Specifically WM isonly adjacent to the GM and can be easily distinguished
from its surroundings so the WM segmentation DRs areapproximate to each other And the visible results of WMare similar in appearance except the fact that HOG-basedmethod mistakenly introduced a small fraction of CSFinto the results It has some difficulty on GM and CSFsegmentation due to the complex anatomical structures andlow contrast HOG and intensity based methods introducemore surroundingnon-GMornon-CSF tissue into theROI ofGM or CSF respectively thus they produce more defects andfuzzy boundaries for different tissues In contrast theGMandCSF tissues generated by ourmethod can be clearly identifiedwith a certain ROI and distinct contours
We then applied our proposed model to segment all422 brain cryosection images Typical images in coronal andsagittal viewpoints and their corresponding segmentationresults (WM GM and CSF) are shown in Figures 7 and 8respectively It is remarkably seen that the results of WMand GM change continuously and their morphological dis-tributions are shown clearly most tiny anatomical structuresare well reserved It is also noticed that the results of CSFseem incomplete and not distributed uniformly This fact isdetermined by the characteristics of the cryosection imagesThese images have such high spatial resolution (0167mmper pixel and 025mm per slice) that it can express finestructures of tissues But since these images were collectedfrom a cadaver the CSF in the brain no longer flowed invivo Due to the effect of gravity the CSF will gather to someplaces rather than uniformly distributing around the surfaceof brain as in live statusThese factors cause the discontinuousdistribution of the CSF segmentation results
BioMed Research International 9
(a)
(b)
(c)
(d)
Figure 7 Coronal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
(a)
(b)
(c)
(d)
Figure 8 Sagittal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
10 BioMed Research International
Figure 9 Three-dimensional surface-rendering reconstruction results of WM GM and CSF based on our segmented images
33 Three-Dimensional Reconstruction Results The CVHcryosection slices are 0167mm per pixel and 025mm thickand such high resolution is very helpful for displaying subtleanatomical structures of brain tissues For a more in-depthunderstanding of these tissues the segmented white mattergray matter and cerebrospinal fluid images are reconstructedusingmarching cubes algorithm to produce 3D surfacemeshThe reconstruction results in different views are shown inFigure 9 From the surface-rendering reconstruction resultsit is seen that the surface of WM is smooth and its sulci andfissures are clearly displayed The distribution of GM shapeis also noticed but the surface of GM reconstruction resultsdoes not look very smooth The reason for this lies in thefact that the GM and CSF are mixed together because of theice crystals in the frozen brain slices so the segmentation ofGM is influenced by its surrounding CSF In spite of it thesulci and cerebral cisterns are also easy to be recognized in3D reconstructed WM and GM
Benefitting from the increasing development of the 3Dreconstruction technology 3D MRI and PET have nowbeen used in clinic and researches But because of theresolution limitation and complexity of brain structures ofthe 2D radiological images (such as CT and MRI) the 3Dreconstruction results are usually unsatisfactory and are hardfor the guidance of clinic operation For the work in ourpaper we focus on the segmentation and reconstruction ofthree kinds of CVHbrain tissuesTheCVHbrain images havehigh spatial resolution of 0167mmper pixel and 025mmperslice after segmentation by the proposed method the high-quality and high-accuracy 2D brain tissues can get well 3Dreconstruction results Some potential applications of the 3Dreconstruction result include the following
(1) The 3D results can be viewed in any orientationbesides the common coronal sagittal and transverseorientations These 3D models (especially for WM
BioMed Research International 11
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
Figure 6 Comparison of the segmentation results with the manually generated segmentation on a cryosection image in CVH dataset (a)shows the original cryosection and its B-channel (in RGB color space) and V-channel (in HSV color space) image (b) shows the manualresults (CSF GM and WM) (c)ndash(g) show the results by the features of our SAE deep learning Intensity PCA HOG and AE respectively
Table 3 Mean and standard deviation of Dice ratio (in ) for thesegmentations obtained by five feature representation methods
the best average Dice ratio of 9069 plusmn 214 (CSF) 9124plusmn 201 (GM) and 9612 plusmn 123 (WM) in a leave-one-outevaluation manner These results have illustrated the strongdiscriminative power of the deep-learned features in braintissues segmentation task
To further demonstrate the advantages of our proposedmodel we visually examine the segmentation results on onecryosection image which is shown in Figure 6 (a) shows theoriginal RGB cryosection image and its B- and V-channelimagesThe ground truth that segmented by experts is shownin (b) (c)ndash(g) present segmentation results of four methodsbased on deep-learned features intensity features PCAfeatures HOG features and AE feautres respectivelyWe cansee that the segmentation results of the proposed model arequite close to the ground truth In contrast other resultseither generate much oversegmentation or fail to segmenttiny anatomical structures accurately Specifically WM isonly adjacent to the GM and can be easily distinguished
from its surroundings so the WM segmentation DRs areapproximate to each other And the visible results of WMare similar in appearance except the fact that HOG-basedmethod mistakenly introduced a small fraction of CSFinto the results It has some difficulty on GM and CSFsegmentation due to the complex anatomical structures andlow contrast HOG and intensity based methods introducemore surroundingnon-GMornon-CSF tissue into theROI ofGM or CSF respectively thus they produce more defects andfuzzy boundaries for different tissues In contrast theGMandCSF tissues generated by ourmethod can be clearly identifiedwith a certain ROI and distinct contours
We then applied our proposed model to segment all422 brain cryosection images Typical images in coronal andsagittal viewpoints and their corresponding segmentationresults (WM GM and CSF) are shown in Figures 7 and 8respectively It is remarkably seen that the results of WMand GM change continuously and their morphological dis-tributions are shown clearly most tiny anatomical structuresare well reserved It is also noticed that the results of CSFseem incomplete and not distributed uniformly This fact isdetermined by the characteristics of the cryosection imagesThese images have such high spatial resolution (0167mmper pixel and 025mm per slice) that it can express finestructures of tissues But since these images were collectedfrom a cadaver the CSF in the brain no longer flowed invivo Due to the effect of gravity the CSF will gather to someplaces rather than uniformly distributing around the surfaceof brain as in live statusThese factors cause the discontinuousdistribution of the CSF segmentation results
BioMed Research International 9
(a)
(b)
(c)
(d)
Figure 7 Coronal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
(a)
(b)
(c)
(d)
Figure 8 Sagittal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
10 BioMed Research International
Figure 9 Three-dimensional surface-rendering reconstruction results of WM GM and CSF based on our segmented images
33 Three-Dimensional Reconstruction Results The CVHcryosection slices are 0167mm per pixel and 025mm thickand such high resolution is very helpful for displaying subtleanatomical structures of brain tissues For a more in-depthunderstanding of these tissues the segmented white mattergray matter and cerebrospinal fluid images are reconstructedusingmarching cubes algorithm to produce 3D surfacemeshThe reconstruction results in different views are shown inFigure 9 From the surface-rendering reconstruction resultsit is seen that the surface of WM is smooth and its sulci andfissures are clearly displayed The distribution of GM shapeis also noticed but the surface of GM reconstruction resultsdoes not look very smooth The reason for this lies in thefact that the GM and CSF are mixed together because of theice crystals in the frozen brain slices so the segmentation ofGM is influenced by its surrounding CSF In spite of it thesulci and cerebral cisterns are also easy to be recognized in3D reconstructed WM and GM
Benefitting from the increasing development of the 3Dreconstruction technology 3D MRI and PET have nowbeen used in clinic and researches But because of theresolution limitation and complexity of brain structures ofthe 2D radiological images (such as CT and MRI) the 3Dreconstruction results are usually unsatisfactory and are hardfor the guidance of clinic operation For the work in ourpaper we focus on the segmentation and reconstruction ofthree kinds of CVHbrain tissuesTheCVHbrain images havehigh spatial resolution of 0167mmper pixel and 025mmperslice after segmentation by the proposed method the high-quality and high-accuracy 2D brain tissues can get well 3Dreconstruction results Some potential applications of the 3Dreconstruction result include the following
(1) The 3D results can be viewed in any orientationbesides the common coronal sagittal and transverseorientations These 3D models (especially for WM
BioMed Research International 11
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
Figure 6 Comparison of the segmentation results with the manually generated segmentation on a cryosection image in CVH dataset (a)shows the original cryosection and its B-channel (in RGB color space) and V-channel (in HSV color space) image (b) shows the manualresults (CSF GM and WM) (c)ndash(g) show the results by the features of our SAE deep learning Intensity PCA HOG and AE respectively
Table 3 Mean and standard deviation of Dice ratio (in ) for thesegmentations obtained by five feature representation methods
the best average Dice ratio of 9069 plusmn 214 (CSF) 9124plusmn 201 (GM) and 9612 plusmn 123 (WM) in a leave-one-outevaluation manner These results have illustrated the strongdiscriminative power of the deep-learned features in braintissues segmentation task
To further demonstrate the advantages of our proposedmodel we visually examine the segmentation results on onecryosection image which is shown in Figure 6 (a) shows theoriginal RGB cryosection image and its B- and V-channelimagesThe ground truth that segmented by experts is shownin (b) (c)ndash(g) present segmentation results of four methodsbased on deep-learned features intensity features PCAfeatures HOG features and AE feautres respectivelyWe cansee that the segmentation results of the proposed model arequite close to the ground truth In contrast other resultseither generate much oversegmentation or fail to segmenttiny anatomical structures accurately Specifically WM isonly adjacent to the GM and can be easily distinguished
from its surroundings so the WM segmentation DRs areapproximate to each other And the visible results of WMare similar in appearance except the fact that HOG-basedmethod mistakenly introduced a small fraction of CSFinto the results It has some difficulty on GM and CSFsegmentation due to the complex anatomical structures andlow contrast HOG and intensity based methods introducemore surroundingnon-GMornon-CSF tissue into theROI ofGM or CSF respectively thus they produce more defects andfuzzy boundaries for different tissues In contrast theGMandCSF tissues generated by ourmethod can be clearly identifiedwith a certain ROI and distinct contours
We then applied our proposed model to segment all422 brain cryosection images Typical images in coronal andsagittal viewpoints and their corresponding segmentationresults (WM GM and CSF) are shown in Figures 7 and 8respectively It is remarkably seen that the results of WMand GM change continuously and their morphological dis-tributions are shown clearly most tiny anatomical structuresare well reserved It is also noticed that the results of CSFseem incomplete and not distributed uniformly This fact isdetermined by the characteristics of the cryosection imagesThese images have such high spatial resolution (0167mmper pixel and 025mm per slice) that it can express finestructures of tissues But since these images were collectedfrom a cadaver the CSF in the brain no longer flowed invivo Due to the effect of gravity the CSF will gather to someplaces rather than uniformly distributing around the surfaceof brain as in live statusThese factors cause the discontinuousdistribution of the CSF segmentation results
BioMed Research International 9
(a)
(b)
(c)
(d)
Figure 7 Coronal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
(a)
(b)
(c)
(d)
Figure 8 Sagittal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
10 BioMed Research International
Figure 9 Three-dimensional surface-rendering reconstruction results of WM GM and CSF based on our segmented images
33 Three-Dimensional Reconstruction Results The CVHcryosection slices are 0167mm per pixel and 025mm thickand such high resolution is very helpful for displaying subtleanatomical structures of brain tissues For a more in-depthunderstanding of these tissues the segmented white mattergray matter and cerebrospinal fluid images are reconstructedusingmarching cubes algorithm to produce 3D surfacemeshThe reconstruction results in different views are shown inFigure 9 From the surface-rendering reconstruction resultsit is seen that the surface of WM is smooth and its sulci andfissures are clearly displayed The distribution of GM shapeis also noticed but the surface of GM reconstruction resultsdoes not look very smooth The reason for this lies in thefact that the GM and CSF are mixed together because of theice crystals in the frozen brain slices so the segmentation ofGM is influenced by its surrounding CSF In spite of it thesulci and cerebral cisterns are also easy to be recognized in3D reconstructed WM and GM
Benefitting from the increasing development of the 3Dreconstruction technology 3D MRI and PET have nowbeen used in clinic and researches But because of theresolution limitation and complexity of brain structures ofthe 2D radiological images (such as CT and MRI) the 3Dreconstruction results are usually unsatisfactory and are hardfor the guidance of clinic operation For the work in ourpaper we focus on the segmentation and reconstruction ofthree kinds of CVHbrain tissuesTheCVHbrain images havehigh spatial resolution of 0167mmper pixel and 025mmperslice after segmentation by the proposed method the high-quality and high-accuracy 2D brain tissues can get well 3Dreconstruction results Some potential applications of the 3Dreconstruction result include the following
(1) The 3D results can be viewed in any orientationbesides the common coronal sagittal and transverseorientations These 3D models (especially for WM
BioMed Research International 11
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
Figure 7 Coronal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
(a)
(b)
(c)
(d)
Figure 8 Sagittal section images (a) and their corresponding SAE segmentation results of WM GM and CSF ((b) (c) and (d))
10 BioMed Research International
Figure 9 Three-dimensional surface-rendering reconstruction results of WM GM and CSF based on our segmented images
33 Three-Dimensional Reconstruction Results The CVHcryosection slices are 0167mm per pixel and 025mm thickand such high resolution is very helpful for displaying subtleanatomical structures of brain tissues For a more in-depthunderstanding of these tissues the segmented white mattergray matter and cerebrospinal fluid images are reconstructedusingmarching cubes algorithm to produce 3D surfacemeshThe reconstruction results in different views are shown inFigure 9 From the surface-rendering reconstruction resultsit is seen that the surface of WM is smooth and its sulci andfissures are clearly displayed The distribution of GM shapeis also noticed but the surface of GM reconstruction resultsdoes not look very smooth The reason for this lies in thefact that the GM and CSF are mixed together because of theice crystals in the frozen brain slices so the segmentation ofGM is influenced by its surrounding CSF In spite of it thesulci and cerebral cisterns are also easy to be recognized in3D reconstructed WM and GM
Benefitting from the increasing development of the 3Dreconstruction technology 3D MRI and PET have nowbeen used in clinic and researches But because of theresolution limitation and complexity of brain structures ofthe 2D radiological images (such as CT and MRI) the 3Dreconstruction results are usually unsatisfactory and are hardfor the guidance of clinic operation For the work in ourpaper we focus on the segmentation and reconstruction ofthree kinds of CVHbrain tissuesTheCVHbrain images havehigh spatial resolution of 0167mmper pixel and 025mmperslice after segmentation by the proposed method the high-quality and high-accuracy 2D brain tissues can get well 3Dreconstruction results Some potential applications of the 3Dreconstruction result include the following
(1) The 3D results can be viewed in any orientationbesides the common coronal sagittal and transverseorientations These 3D models (especially for WM
BioMed Research International 11
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
Figure 9 Three-dimensional surface-rendering reconstruction results of WM GM and CSF based on our segmented images
33 Three-Dimensional Reconstruction Results The CVHcryosection slices are 0167mm per pixel and 025mm thickand such high resolution is very helpful for displaying subtleanatomical structures of brain tissues For a more in-depthunderstanding of these tissues the segmented white mattergray matter and cerebrospinal fluid images are reconstructedusingmarching cubes algorithm to produce 3D surfacemeshThe reconstruction results in different views are shown inFigure 9 From the surface-rendering reconstruction resultsit is seen that the surface of WM is smooth and its sulci andfissures are clearly displayed The distribution of GM shapeis also noticed but the surface of GM reconstruction resultsdoes not look very smooth The reason for this lies in thefact that the GM and CSF are mixed together because of theice crystals in the frozen brain slices so the segmentation ofGM is influenced by its surrounding CSF In spite of it thesulci and cerebral cisterns are also easy to be recognized in3D reconstructed WM and GM
Benefitting from the increasing development of the 3Dreconstruction technology 3D MRI and PET have nowbeen used in clinic and researches But because of theresolution limitation and complexity of brain structures ofthe 2D radiological images (such as CT and MRI) the 3Dreconstruction results are usually unsatisfactory and are hardfor the guidance of clinic operation For the work in ourpaper we focus on the segmentation and reconstruction ofthree kinds of CVHbrain tissuesTheCVHbrain images havehigh spatial resolution of 0167mmper pixel and 025mmperslice after segmentation by the proposed method the high-quality and high-accuracy 2D brain tissues can get well 3Dreconstruction results Some potential applications of the 3Dreconstruction result include the following
(1) The 3D results can be viewed in any orientationbesides the common coronal sagittal and transverseorientations These 3D models (especially for WM
BioMed Research International 11
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
and GM) can help obtain the anatomical knowledgeof 3D structures and their adjacent relationship inspace In addition we can identify the structures bycomparing radiological image with the anatomicalimage
(2) The 3D result in our work is applicable for teachingsectional anatomy since there is rarely direct-viewingmodel that canmake the understanding of anatomicalstructures easier The varying viewpoints of the 3Dmodel are helpful for observing the tiny structuresin specific positions of human brain and medicalstudents only need tomove theirmouse to control theperspective of displaying
4 Conclusion
We have presented a supervised learning-based CVH braintissue segmentation method using the deep-learned featureswith multilayer SAE neural network The major differencebetween our proposed feature extraction method and con-ventional hand-crafted features such as intensity PCA andHistogram of Gradient (HOG) is that it can dynamicallylearn the most informative features adapted to the inputdataset at hand The discriminative ability of our proposedmodel is evaluated and compared with other types of imagefeatures Experimental results validate the effectiveness ofour proposed method and show that it significantly outper-forms the methods based on typical hand-crafted featuresIn addition the high-resolution 3D tissue surface meshesare reconstructed based on the segmentation results byour method with the resolution of 0167mm per pixel and025mm per slice much more tiny than the 3 T (even 7 T)MRI brain images Furthermore since the procedure offeatures extraction by ourmethod is independent of the CVHdataset our method can be easily extended to segment othermedical images such as cell images and skin cancer imagesCVH dataset contains serial transverse section images of thewhole human body which is large in volume Pure manual orsemiautomatic segmentation of those images is quite time-consuming so a large proportion of data still remain to beexploitedThough the work in our paper only had segmentedtheWMGM and CSF of the brain tissue it actually providesa reference for automatically or semiautomatically processingsuch real-color and high-resolution images
Recent studies show that neural network can yield morepromising performance on image recognition task withdeeper hidden layers [35 36] we will explore parallel SAEwith more hidden layers as well as more training data in thefuture In addition the number of neural units in each hiddenlayer may affect the segmentation performance in a certaindegree We will further investigate the influence of hiddenneural units to segmentation performance Furthermore themodel uses a classical Softmax classifier to predict labels of theinput patches and we will consider the influence of differentclassifiers in the future research
Conflict of Interests
The authors declare that they have no conflict of interests
Acknowledgments
This work was partially supported by the Chinese NationalScience Foundation (NSFC 60903142 and 61190122) theChina Postdoctoral Science Foundation (2013T608412012M521677 and Xm201306) and Fundamental ResearchFunds for the Central Universities (106112015CDJXY120003)
References
[1] Y Wu L-W Tan Y Li et al ldquoCreation of a female and malesegmentation dataset based on chinese visible human (CVH)rdquoComputerized Medical Imaging and Graphics vol 36 no 4 pp336ndash342 2012
[2] S-X Zhang P-A Heng and Z-J Liu ldquoChinese visible humanprojectrdquo Clinical Anatomy vol 19 no 3 pp 204ndash215 2006
[3] M Li X-L ZhengH-Y Luo et al ldquoAutomated segmentation ofbrain tissue and white matter in cryosection images from Chi-nese visible human datasetrdquo Journal of Medical and BiologicalEngineering vol 34 no 2 pp 178ndash187 2014
[4] Q Mahmood A Chodorowski and M Persson ldquoAutomatedMRI brain tissue segmentation based on mean shift and fuzzyc-means using a priori tissue probability mapsrdquo IRBM vol 36no 3 pp 185ndash196 2015
[5] G Paul T Varghese K Purushothaman and N A Singh ldquoAfuzzy c mean clustering algorithm for automated segmentationof brain MRIrdquo in Proceedings of the International Conferenceon Frontiers of Intelligent Computing Theory and Applications(FICTA) 2013 vol 247 of Advances in Intelligent Systems andComputing pp 59ndash65 Springer Berlin Germany 2014
[6] M A Balafar ldquoGaussian mixture model based segmentationmethods for brain MRI imagesrdquo Artificial Intelligence Reviewvol 41 no 3 pp 429ndash439 2014
[7] P Anbeek I Isgum B J M van Kooij et al ldquoAutomaticsegmentation of eight tissue classes in neonatal brain MRIrdquoPLoS ONE vol 8 no 12 Article ID e81895 2013
[8] P Moeskops M A Viergever M J N L Benders and IIsgum ldquoEvaluation of an automatic brain segmentationmethoddeveloped for neonates on adult MR brain imagesrdquo in MedicalImaging Image Processing vol 9413 of Proceedings of SPIE pp1ndash6 March 2015
[9] S J Hussain T S Savithri and P S Devi ldquoSegmentation oftissues in brain MRI images using dynamic neuro-fuzzy tech-niquerdquo International Journal of Soft Computing and Engineeringvol 1 no 6 pp 2231ndash2307 2012
[10] A Demirhan M Toru and I Guler ldquoSegmentation of tumorand edema along with healthy tissues of brain using waveletsand neural networksrdquo IEEE Journal of Biomedical and HealthInformatics vol 19 no 4 pp 1451ndash1458 2015
[11] S Liao Y Gao A Oto and D Shen ldquoRepresentation learninga unified deep learning framework for automatic prostate MRsegmentationrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2013 vol 8150 of Lecture Notesin Computer Science pp 254ndash261 Springer Berlin Germany2013
[12] N Nabizadeh N John and C Wright ldquoHistogram-basedgravitational optimization algorithm on single MR modalityfor automatic brain lesion detection and segmentationrdquo ExpertSystems with Applications vol 41 no 17 pp 7820ndash7836 2014
[13] M Jirik T Ryba and M Zelezny ldquoTexture based segmentationusing graph cut and Gabor filtersrdquo Pattern Recognition andImage Analysis vol 21 no 2 pp 258ndash261 2011
12 BioMed Research International
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009
[14] A P Nanthagopal and R Sukanesh ldquoWavelet statistical texturefeatures-based segmentation and classification of brain com-puted tomography imagesrdquo IET Image Processing vol 7 no 1pp 25ndash32 2013
[15] W X Chen and X Lin ldquoBig data deep learning challenges andperspectivesrdquo IEEE Access vol 2 pp 514ndash525 2014
[16] A Krizhevsky I Sutskever G E Hinton and A KrizhevskyldquoImagenet classification with deep convolutional neural net-worksrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo12) pp 1097ndash1105 December 2012
[17] Y Bengio ldquoLearning deep architectures for AIrdquo Foundationsand Trends in Machine Learning vol 2 no 1 pp 1ndash127 2009
[18] G E Hinton S Osindero and Y-W Teh ldquoA fast learningalgorithm for deep belief netsrdquoNeural Computation vol 18 no7 pp 1527ndash1554 2006
[19] YGuoGWu LA Commander et al ldquoSegmenting hippocam-pus from infant brains by sparse patch matching with deep-learned featuresrdquo in Medical Image Computing and Computer-Assisted InterventionmdashMICCAI 2014 pp 308ndash315 Springer2014
[20] G E Hinton and R R Salakhutdinov ldquoReducing the dimen-sionality of data with neural networksrdquo Science vol 313 no5786 pp 504ndash507 2006
[21] A Krizhevsky andG E Hinton ldquoUsing very deep autoencodersfor content-based image retrievalrdquo in Proceedings of the 19thEuropean Symposium on Artificial Neural Networks (ESANNrsquo11) pp 489ndash494 Bruges Belgium April 2011
[22] W Zhang R Li H Deng et al ldquoDeep convolutional neuralnetworks for multi-modality isointense infant brain imagesegmentationrdquo NeuroImage vol 108 pp 214ndash224 2015
[23] J Dai K He and J Sun ldquoConvolutional feature masking forjoint object and stuff segmentationrdquo httparxivorgabs14121283
[24] F Agostinelli M R Anderson and H Lee ldquoAdaptive multi-column deep neural networks with application to robust imagedenoisingrdquo in Proceedings of the Advances in Neural InformationProcessing Systems (NIPS rsquo13) pp 1493ndash1501 Stateline NevUSA December 2013
[25] J Xie L Xu and E Chen ldquoImage denoising and inpaintingwith deep neural networksrdquo in Advances in Neural InformationProcessing Systems pp 341ndash349 MIT Press 2012
[26] H-C Shin M R Orton D J Collins S J Doran and M OLeach ldquoStacked autoencoders for unsupervised feature learningand multiple organ detection in a pilot study using 4D patientdatardquo IEEE Transactions on Pattern Analysis and MachineIntelligence vol 35 no 8 pp 1930ndash1943 2013
[27] J Xu L XiangQ Liu et al ldquoStacked sparse autoencoder (SSAE)for nuclei detection on breast cancer histopathology imagesrdquoIEEE Transactions on Medical Imaging 2015
[28] S-X Zhang P-AHeng Z-J Liu et al ldquoCreation of the Chinesevisible human data setrdquoTheAnatomical Record Part BThe NewAnatomist vol 275 no 1 pp 190ndash195 2003
[29] Y Qi Y Wang X Zheng and Z Wu ldquoRobust feature learningby stacked autoencoder with maximum correntropy criterionrdquoin Proceedings of the IEEE International Conference onAcousticsSpeech and Signal Processing (ICASSP rsquo14) pp 6716ndash6720Florence Italy May 2014
[30] T Amaral L M Silva L A Alexandre C Kandaswamy JM Santos and J M de Sa ldquoUsing different cost functions totrain stacked auto-encodersrdquo in Proceedings of the 12th MexicanInternational Conference on Artificial Intelligence (MICAI rsquo13)
F Castro A Gelbukh and M G Mendoza Eds pp 114ndash120IEEE Mexico City Mexico November 2013
[31] A Coates A Y Ng and H Lee ldquoAn analysis of single-layernetworks in unsupervised feature learningrdquo in Proceedings ofthe 14th International Conference on Artificial Intelligence andStatistics (AISTATS rsquo11) pp 215ndash223 Fort Lauderdale Fla USAApril 2011
[32] J Shlens A Tutorial on Principal Component Analysis SystemsNeurobiology Laboratory 2005
[33] Y Zhang and L Wu ldquoAn MR brain images classifier via prin-cipal component analysis and kernel support vector machinerdquoProgress in Electromagnetics Research vol 130 pp 369ndash3882012
[34] O Deniz G Bueno J Salido and F De la Torre ldquoFacerecognition using histograms of oriented gradientsrdquo PatternRecognition Letters vol 32 no 12 pp 1598ndash1603 2011
[35] S Gao Y Zhang K Jia J Lu and Y Zhang ldquoSingle sample facerecognition via learning deep supervised autoencodersrdquo IEEETransactions on Information Forensics and Security vol 10 no10 pp 2108ndash2118 2015
[36] K Jarrett K Kavukcuoglu M Ranzato and Y LeCun ldquoWhatis the best multi-stage architecture for object recognitionrdquo inProceedings of the 12th International Conference on ComputerVision (ICCV rsquo09) pp 2146ndash2153 IEEE Kyoto Japan October2009