Automatic Lumbar Vertebrae Segmentation from T1 …Automatic Lumbar Vertebrae Segmentation from T1-Weighted Images using Clustering Approaches: A Comparative Study ... The T1 - Weighted

Automatic Lumbar Vertebrae Segmentation from T1-Weighted Images

using Clustering Approaches: A Comparative Study

1J.V.Shinde,

2Dr. Y.V.Joshi,

2Dr. R.R.Manthalkar

1S.G.G.S. Institute of Engineering &Technology, Nanded, India

1jv.shinde@rediffmail.com

2S.G.G.S. Institute of Engineering &Technology, Nanded, India

2yvjoshi@sggs.ac.in,

2rrmanthalkar@sggs.ac.in

Abstract For clinicians assessment and correct diagnosis of lower back pain is very challenging task

due to wide variations in shape, size, position and orientation of spinal objects like vertebrae and

intervertebral discs. Degenerative lumbar spine diseases comprises of disc herniation and Modic

changes which are directly associated with the lower back pain and very common in aged people.

Modic changes and different types of disc herniation are visible on Magnetic Resonance Imaging

(MRI) scans.In particularmild to severe degenerations are visible the form of signal intensity

variations. Degenerative changes are assessed by comparing with normal signalintensities. The T1-

Weighted MR images are recommended in assessment of Modic changes while, the T2-weighted

images are studied for IVD degeneration.For better characterization of such changes accurate

segmentation is essential crucial step.In this paper we have presented a performance analysis of

popular clustering methodsutilized for delineating spinal objects namely Otsu, K-means and Fuzzy c-

means with spatial constraints. Results of these methods are subsequentlybenchmarked with the state-

of-the art method called marker controlled watershed transform which is popular technique for

segmentation of medical images. In the proposed work it is primarily used to check consistency in

case ofcomplex spine structures exists due to severe degeneration and presence of different types of

artefacts found in MRI scans. We have experimented with total 106 T1-weighted images acquired

from 106 subjects. Comparative analysis is performed fordifferent performance measures metrics viz.-

DSC,Jaccard Index, Sensitivity and Specificity. Experimental results shows that marker controlled

watershed method outperform clustering approaches for spinal objects segmentation in severe

deformity and degeneration cases.

.

Keywords: Spine MRI segmentation, Clustering, Marker controlled watershed

transform.

1.Introduction

Survey shows that 80% population in the world experiences back pain at some time in their

lifespan. Unhealthy lifestyle, bad postures and sedentary jobs causes lots of stress which developsback

pain [1][2][3].There are various spine related diseases and abnormalities found in children to adults.

In children a deformed spine with an S shaped curvature is observed. While degenerative changes

appear amongst aged people.Modic changes [4] are terms which are used to describe the changes of

the vertebral endplate and marrow. They are related with spinal degeneration surrounding the

intervertebral disc,and are most common in the lumbar spine.In addition, it is noted that they can

occur at any level. They are most commonly observed at L4-L5 and L5-S1 locations, where the

changes occur are variable too. Generally they found anteriorly, but there are also cases where almost

complete vertebral involvement is observed. Lumbar disc degenerationsare range from mild to severe

termed as herniation, protrusion, sequestration.Whole spinal column is divided into 5 sections namely

cervical, thoracic, lumbar, sacrum and coccyx.The cervical section contains 7 vertebrae C1-C7 and

discs, Thoracic contains 12 vertebrae T1-T12 and discs, lumbar section contains 5 vertebrae L1-L5

and 5 discs, sacrum shows fused vertebrae S1-S5 which articulates with the hip bone of pelvis and

International Journal of Pure and Applied MathematicsVolume 119 No. 12 2018, 12923-12935ISSN: 1314-3395 (on-line version)url: http://www.ijpam.euSpecial Issue ijpam.eu

12923

finally entire column is terminated by tiny coccyx section shown in Fig. 1. MRI scans are captured in

three planes axial/transverse, sagittal, and coronal. Both sagittal and transverse images are valuable

for examining the spinal cord and structural deformation such as disc herniation. An axial image

depicts excellent visualization of nerve roots and possible disc fragments. Scanning planes are

illustrated in Fig.2.

Figure 1. (a) Whole spine curve (b) Cervical curve (c) Thoracic curve (d) Lumbar-sacrum-coccyx curves.

Figure 2 Three planes of spine MR Imaging (a) Axial (b) Sagittal (c) Coronal White arrow points to vertebra

and Red arrowshows intervertebral disc location in all three planes

MRI allows more clear visualization of all soft tissues and has better capability to

differentiate between normal and abnormal tissues [5]. Degeneration changes are shown on MRI in

terms of signal intensity changes whether it is to be fat, oedema or sclerosis. On MRI Modic

changes (1-3) [4][5][6] will be detected by observing signal alterations in the endplates parallel to

the disc.MRI shows only shades of grey from bright to dark. When comparing T1-weighted with

T2-Weighted images, on T1-Weighted scans, the fluid is dark, and on T2-weighted scans fluid is

bright. Cortical bone and spinal cord matter are represented in darker grey intensity on both

images.Subcutaneous fat hasbrighter grey intensity on both scans. The main difference

capturedwith MRI is the brightness of grey level which fluid will reflect. Fluids make the clinicians

easier to observe as one compares the images.T1& T2-weighted imagesare widelyused with Fast

Spin Echo (FSE) technique. Sometimes a combination of T1 and T2 weighted images is essential.

Figure 3 shows sample T1 & T2 weighted scans.

A difficulty with consistentautomated segmentation isdue to enormous variations in backbone

structure like curvature of spinal column, shape, size, appearances of discs and vertebraeas depicted

in Figure 4. Intensity inhomogeneity due to B0 & B1 field, presence of artefacts, low resolution

images misleads the segmentation algorithms.

International Journal of Pure and Applied Mathematics Special Issue

12924

Figure 3 (A) T1-Weighted Image (B) T2-Weighted Image. arrows are showing vertebral bodies

Figure 4.AnatomicalVariability and various vertebra disorders in our dataset

1.1.Spine Degeneration :Clinical Context

As indicated in Fig.5, Vertebra degeneration which is observed in three types (I-III)referred as Modic

changes. Type-I changes are acute degenerative changes represented with low signal intensity in T1

and high signal intensity in T2 image Type-II changes can be distinguished with increased signal

intensity in both T1 and T2 image. Type-III changes are found by locating low signal intensity on

both T1 and T2 weighted images[4][6].

Figure 5. (a) T1-Weighted images (b) T2-weighted images both are showing vertebra degeneration named

as Modic changes.

a b a b

International Journal of Pure and Applied Mathematics Special Issue

12925

1.2 Intervertebral Disc Degenerations

Intervertebral degeneration is primary cause of lower back pain resulting in compression of spinal

nerves and adjacent vertebrae. Intervertebral discs acts as a shock absorber and provides flexibility.

Natural aging, genetic, biological and environmental factors may play important role in degeneration.

There are two anatomical parts of IVD; center part is gelatinous known as„Nucleus Pulposus‟ and

outer wall is „Annulus Fibrosus‟.Degeneration categories are found and classified with Pfirrmann‟s

grading method by examining the parameters like the signal intensity on MRI, height of disc etc. [7].

Figure 6. Degenerated discs from lumbar curve are shown by arrows on T2-Weighted Scans.

2. Related Works

In spine degeneration analysis it is required to partition the image into different region of interest

like degeneratedvertebral bodies and intervertebral discs. Various segmentation methods have been

proposed in literature in image processing and analysis domain.There are some challenges in

consistent segmentation of spinal objects using MRI scans. First major hurdle isa presence of internal

and external artefacts and sensitivity towards spurious variations of image intensities. MR image

intensities of the same subject are not constant. Rather they vary with pulse sequences parameters,

different implementation of protocol sequences. Last but not least, due to low and strongly anisotropic

voxel resolution segmentation is not consistent. Few popular segmentation approaches used in MR

image analysis are namely thresholding[8], clustering, graph cut [9], atlasbased [10], active contours,

edge based[11] , and watershed transform[12][13].However, every approach has its own lacunas.

Thresholding is the simplest approach used in image analysis. Global, local, clustering based,

histogram shape based, entropy based are some of its variants. Otsu‟salgorithm [14] is a clustering

based method which is searching for an optimum threshold that minimizes intra-class variance which

is a weighted sum of variances of the two classes. In parallel it maximizes inter-class variance.

However, it exhibits good performance if the histogram of image is having bimodal distribution. Total

variance is given by equation 1.

2

2111

22 )]()()][(1)[()( tttqtqtw (1)

Where, )(2 tw isintra-class and

2

2111 )]()()][(1)[( tttqtq = 𝜎𝐵2(t) is inter-class variance.

K-means [15] and Fuzzy c-means [16][17] are clustering techniques which partition image into

groups. Clustering is achieved by iteratively minimizing a cost function that depends on the distance

International Journal of Pure and Applied Mathematics Special Issue

12926

of pixels to the cluster centres in the feature space.K-means is an iterative process that

assigns n observations to exactly one of k clusters defined by centroids which are randomly chosen;

where k value is selected before the algorithm starts. But it is strongly affected by outliersandit slows

with increase in dimensionality. Algorithmic steps are as follows

1. Randomly initialize k number of clusters and their corresponding cluster centers as 𝐶𝑘 .

2. Compute using Euclidean distance between all centers𝐶𝑘 . And each data point 𝑄 𝑥, 𝑦 as per

eq.2 and assign to closet center

𝐷 = 𝑄 𝑥, 𝑦 − 𝐶𝑘 2 (2)

3. Update cluster center using eq. 3

𝑪𝒌 = 𝒙∈𝒌 𝒚∈𝒌 𝑸(𝒙,𝒚)

𝒏∈𝒌 (3)

4. Repeat step 3 and 4 until it meets the convergence criteria.

Fuzzy c-means (FCM) algorithm providesbetter results than k-means in case of overlapped

dataset. It allows one sample of data belongs to two or more clusters with varying degree of

membership. Conventional FCM does not fully utilize the spatial information into membership

function & its objective function is not considering any spatial dependence among the pixels of

image. Secondly, the membership function of FCM is mostly decided by Euclidean distance which

measures the similarity between the pixel intensity and cluster centre. Higher membership depends

upon closer intensity values to the cluster centre. Fuzzy c means algorithm is highly affected by noise.

Fuzzy c means with addition of spatial information in membership function is proposed by [18].It is

iterative optimization that minimizes the cost function. sFCMassumes that neighboring pixels possess

similar feature value. It is a spatial relationship function. Effect of noise is also considerably less.

ℎ𝑖𝑗 = 𝑢𝑖𝑘𝑘∈𝑁𝐵(𝑥𝑗 ) (4)

Where 𝑁𝐵(𝑥𝑗 ) is square window centered pixel 𝑥𝑗 in the spatial domain, Window size is 5X5 spatial

function is incorporated in membership function, 𝑝 and 𝑞 are used to control relative importance of

both functions. Where 𝝁𝒊𝒋 represent membership of pixel 𝒙𝒋 in the 𝑖thcluster, fuzziness=2.

𝑢𝑖𝑗′ =

𝑢𝑖𝑗 𝑝

ℎ𝑖𝑗𝑞

𝑢𝑘𝑗 𝑝

ℎ𝑘𝑗𝑞 𝑐

𝑘=1

(5)

In watershed transform,watershed refers to a ridge that divides areas by different river

system.It is widely used in medical image analysis because it can easily cope with anatomical

variability of shapes and topologies. This method has been used with other state of the art methods for

spine image segmentation but rarely used to handle more complex spine disorders. It detects ridges

and valleys from image where image is viewed as a topological surface and altitude is represented by

intensity of the pixel. Catchment basins represent regions of homogeneous intensity,

butoversegmented image is obtained due to noise or other unwanted structuresthat appears in gradient

image. Themarker controlledwatershed transform method modifies the gradient image by using

foreground and background marker. Markers will keep only most significant and relevant contours for

structure of interest. Markers are created with morphological operations like erode, dilate open etc.

Square and Rectangle structuring elements are suitable for vertebra & intervertebral disc [19].

Algorithmic steps are as follows.

1. Compute gradient magnitude as the segmentation function using edge operator

2. Spot foreground markers using morphological technique called open by reconstruction to

create flat maxima inside each object.

International Journal of Pure and Applied Mathematics Special Issue

12927

3. Computer background markers. A dark pixel belongs to the background so we can use

thresholding operation followed by distance transform of thresholded image and then looking

for watershed ridge lines.

4. Computation of watershed transform of segmentation function i.e. gradient magnitude image

Figure 7. (a) Original vertebra image (b) gradient image(c) distance transform.

3.Proposed work

Initial input to all four methods is the T1-weighted mid-sagittal DICOM image. In order to get the

consistent contrast throughout,the image normalization is done using contraststretching based on

simple linear mapping. Normalized image is further processed by different methods.After applying

different segmentation algorithms we have extracted individual vertebra by applying a series of

morphological operations like hole filling, erosion, dilation using square structuring element. We are

focusing on lumbar vertebrae L1-L5 which are extracted on the basis of aspect ratio which is found to

be in the range of 1 to 2.The implementation is done using MATLAB2017a.

Figure 8. Vertebral bodies segmentation process workflow

International Journal of Pure and Applied Mathematics Special Issue

12928

3.1 Clinical MR Imaging Dataset Details

Mid-sagittal DICOM MR images are acquired from Samarth diagnostic center for 106 individuals of

both sexes with age varying from 20 to 89yrs. Imaging protocols with parametersas given below.

T1-weighted FRFSE SAG TR=700ms, TE=114.34ms, FOV280X280mm, Slice

Thickness=4mm, image matrix size is 512 X512, no. of excitations=2

T2-weighted FRFSE SAG TR=4220ms, TE=105.25ms, FOV300X300mm, Slice

Thickness=4mm, Matrix size is 512X512, no. of excitations=1

Reference data (Ground Truth) are produced by manually tracing the vertebral bodies in the sagittal

T1-weighted images. Manual segmentation of 5 images is done by experienced radiologist while rests

of the images are segmented by first author. Manual segmentation takes 3-4 minutes per vertebra. We

are only focusing on lumbar vertebrae L1-L5.

Figure 9.(a) OriginalT1 image and its (b) Ground Truth: vertebrae from bottom L5-L4-L3-L2-L1

3.2 Experimental Results and Validation

3.2.1 Otsu’s Method

Otsu‟smethodis tested on 106 T1-weighted MR images of spine. Our aim is to segment vertebral

bodies form the background which consists of muscles, discs and nerves etc. The process is shown in

Fig. 10 in stepwise manner.

Figure 10.(a) Originalimage (b) Thresholded image (c) segmentedimage after post processing (d) Overlay image

3.2.2 K-means Method

K-means isan unsupervised clustering method. It is required to provide 𝑘 value which indicates

number of clusters. Image will be divided into 3 clusters or classes if 𝑘 = 3. Figure 11 shows three

clusters.

International Journal of Pure and Applied Mathematics Special Issue

12929

Figure 11.(a) Original image k=3 (b) Cluster 1 (c) Cluster 2 (d) cluster 3 ,for vertebrae segmentation class 2 image will be

processed further.

3.2.3 sFCM Method

Figure12.sFCM processing using 3 clusters followed by final segmented image and labeled image.

International Journal of Pure and Applied Mathematics Special Issue

12930

3.2.4 Marker controlled watershed transform

Figure13 .Marker controlled watershed transform (a) original image (b) gradient image(c) regional maxima (d)

regional maxima superimposed on original image (e) distance transform (f) ridge lines (g) marker image (h)

watershed colour image (i) overlay image showing segmented part in red colour superimposed on original

image.

All clustering methods are applied on all 106 images, but in some cases theyfail to provide acceptable

results. Clustering approaches fails to adapt with the complex spine structures. MR image scansshown

in Fig.14 illustrate shape and size disorders exists due to trauma or degeneration. Marker controlled

watershed method shows better results on complex topology and less sensitive to noise as compared

to clustering methods.

Segmentation results of all above methods are validated using spatial overlap index DSC, Jaccard

Index, Sensitivity and Specificity. Automated segmented image (A) is compared with ground truth

image (R).

𝐷𝑖𝑐𝑒 =2|𝑉 𝐴 ∩ 𝑅 |

|𝑉 𝐴 |+|𝑉 𝑅 | 𝐽𝑎𝑐𝑐𝑎𝑟𝑑 𝑖𝑛𝑑𝑒𝑥 =

𝑡𝑝𝑡𝑝 + 𝑓𝑝 + 𝑓𝑛

(6)

𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =𝑡𝑛

𝑡𝑛 + 𝑓𝑝 𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑦 =

𝑡𝑝𝑡𝑝 + 𝑓𝑛

(7)

International Journal of Pure and Applied Mathematics Special Issue

12931

Figure14. Results on image scans with complex spine structure and non-homogeneous image area and noise

segmented area in red colour superimposed on original image (a) Otsu (b) K-means (c) sFCM (d) marker

controlled watershed transform

Table 1.Performance analysis of Otsu, K-means, sFCM with Marker controlled watershed method.

Evaluation is done only for .L1-L5 vertebrae from lumbar section for 106 subjects

Method Performance validation Metrics

Dice Jaccard Sensitivity Specificity

Otsu thresholding 0.82 0.76 0.80 0.71

K-means Clustering 0.88 0.82 0.84 0.79

sFCM 0.90 0.83 0.91 0.79

Marker controlled watershed 0.92 0.83 0.90 0.81

Figure15. Graph showing comparative performance analysis

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Dice Jaccard Sensitivity Specificity

Otsu Thresholding

K-means

sFCM

Marker Controlled watershed

International Journal of Pure and Applied Mathematics Special Issue

12932

4. Conclusion

In this work, comparative analysis of three popular clustering methods namely Otsu‟s, K-means and

sFCM is carried out for lumbar vertebrae segmentation from spine MR images using our own image

dataset. Otsu‟s method fails if non-homogeneity is found in an image and works better when image

histogram is bimodal. While Clustering methods like K-means and sFCM are susceptible to noise,

outliers and provide poor quality results in case of complex disorders due to trauma or severe

degenerations. In contrast withclustering methods marker controlled watershed method can handle

complex spine structure. Clustering methods are computationally expensive due to calculation of

membership function and iterative process. In addition, increase in number of clusters will require

more execution time. However, marker controlled watershed is inexpensive in terms of time and

computations. In future to get better accuracy we may propose hybrid approach using watershed and

clustering methods. Also we can improve results by using neural networks or deep learning based

methods. This study will be further extended to extract features from vertebral bodies and

intervertebral discs to perform the classification of severity of degeneration and deformity

References

[1] SupreetBindra, Sinha A.G.K. and Benjamin ,.Epidemiology of Low Back Pain in Indian

population: A Review, International Journal of Basic and Applied Medical Sciences ISSN:

2277-2103 (Online) 015 Vol. 5 (1) January-April, pp. 166-179

[2] MohdNazeer, Surender M Rao, SimmiSoni, Ravinder Lower Back Pain in South Indians:

Causative Factors and Preventive Measures ,Sch. J. App. Med. Sci., 2015; 3(1D):234-243

ISSN 2347-954X

[3] David F Fardon, Lumbar disc nomenclature: version 2.0 Recommendations of the combined

task forces of the North American Spine Society, the American Society of Spine Radiology

and the American Society of Neuroradiology Review Article The Spine Journal 14 (2014)

2525–2545.

[4] H. B. Albert, P. Kjaer, T. S. Jensen, J. S. Sorensen, T. Bendix, and C. Manniche, “Modic

changes possible causes and relation to low back pain,” Med. Hypotheses, vol. 70, no. 2, pp.

361–368, 2008

[5] Bezdek J, Hall L, Clarke L. Review of MR image segmentation using pattern recognition.

Med Phys 1993; 20:1033–48.

[6] Amir Jamaludin,TimorKadir,AndrewZisserman, Automated classification and evidence

visualization in Spinal MRI, Medical Image Analysis 41 (2017) 63-73

[7] CWA.Pfirrmann,Alexander Metzdorf, Marco Zanetti, JuergHodler, Magnetic Resonance

Classification of Lumbar Intervertebral Disc Degeneration, SPINE 2001,26(17),1873-8.

[8] P. Sahoo, S. Soltani, and A. Wong, “Survey of thresholding techniques,” Comp. Vis. Graph.

Image Proc., Vol. 41, no. 2, pp. 233_60, 1988.

[9] Huang SH, Chu YH, Lai SH, Novak CL, Learning-based vertebra detection and iterative

normalized-cut segmentation for spinal MRI. IEEE Trans Med Imaging 28:1595–1605,2009

[10] Sofia Michopoulou “Atlas Based Segmentation of Degenerated Lumbar Intervertebral Discs

from MR Images of the Spine” IEEE Trans. Med. Imag.2010

[11] D.L.Pham, C.Y.Xu, J.L.Prince, A survey on current methods in Medical Image

Segmentation in Annual Review of Biomedical Engineer, 2000.

[12] Vincent, L., Soille, P.: Watersheds in digital spaces: An efficient algorithm based on

immersion simulations. IEEE Transactions on Pattern Analysis and Machine Intelligence

13(6), 583–598 (1991) 0162-8828.

[13] Marko Rak,KlausD.Tonnies,On computerized methods for spine analysis in MRI :a

systematic review.,Int J CARS11:1445-1465, 2016.

International Journal of Pure and Applied Mathematics Special Issue

12933

[14] Otsu, N.: A threshold selection method from gray-level histograms. IEEE Transactions on

Systems, Man and Cybernetics SMC-9(1), 62–66 (1979) 0018-9472. Information Sci Div.,

Electrotech. Lab., Tokyo, Japan

[15] NameirakpamDhanachandra, KhumanthemManglem, YambemJinaChanu, Image

segmentation using K-means clustering algorithm and subtractive clustering algorithm. In

international Multi-conference on Information Processing, Procedia Computer Science Vol.54

2015, pp. 764-771

[16] Ahmed MN, Yamany SM, Mohamed N, Farag AA, Moriarty T. A modified fuzzy c-means

algorithm for bias field estimation and segmentation of MRI data. IEEE Trans Med Imaging

21:193–9.2002.

[17] D.L.Pham, J.L.Prince, Adaptive fuzzy segmentation of magnetic resonance images. IEEE

Trans. Med. Imag. Vol.18 pp 737-752 1999.

[18] K. Chuang, H. Tzeng, J. W. S Chen, Fuzzy c-means clustering with spatial information for

image segmentation. Comput. Med. Imag. Graph., Vol. 30, no. 1, pp. 9-15, 2006

[19] Claudia Cheverfils,” Texture analysis for Automatic Segmentation of Intervertebral Disk of

Scoliotic Spines from MR images”, IEEE Trans. Med. Imag.2009

International Journal of Pure and Applied Mathematics Special Issue

12934

12935

12936