Proceedings of the 3rd International Conference on E ...zabulis/EHB11_Retina_Analysis.pdf · 2. Segmentation Given the pre-processed image, the segmentation of blood vessels is performed

An Image Analysis Framework for the Early Assessment of Hypertensive Retinopathy Signs

Georgios C. Manikis1, Vangelis Sakkalis1, Xenophon Zabulis1, Polykarpos Karamaounas1, Areti Triantafyllou2, Stella Douma2, Chrysanthos Zamboulis2, Kostas Marias1

1 Institute of Computer Science, Foundation for Research and Technology, Hellas 2 Hypertension Unit, 2nd Propaedeutic Department of Internal Medicine, Hippokration Hospital, Aristotle University of

Thessaloniki, Hellas {gmanikis; sakkalis; zabulis; karam}@ics.forth.gr, [email protected], {sdouma; zambouli}@med.auth.gr,

[email protected]

Abstract- This paper presents a framework for the detection and measurement of retinal vessels in fundoscopy images. The proposed method segments retinal vessels, enabling the actual measurement of the vessel diameter. A Graphical User Interface (GUI) further supports a number of methods for the automatic and interactive measurement of vessel diameters at any selected point or region of interest. In addition, it provides methods for editing the vessel representation in order to recover from possible segmentation misclassifications. All the above steps are integrated into a clinical application capable of supporting vascular risk stratification in persons with hypertension. The performance of the segmentation is assessed on two publicly available DRIVE and STARE datasets. The experimental results show that the method can achieve a high accuracy on both datasets with very low computational cost, being comparative or slightly better than widely used supervised and unsupervised methods.

Keywords: Retinal images; vessel segmentation; vessel enhancement filtering; vascular measurements; retinal image toolkit

I. INTRODUCTION

The retina provides an open and accessible window for studying the microcirculation in the human body. Retinal vessels can be easily visualized with non-invasive techniques providing valuable information in patients of hypertension and cardiovascular diseases [1]. For many years physicians used to perform traditional fundoscopy, which provides an overview of the retina and vitreous of the eye, by the use of a direct or indirect ophthalmoscope or with a slit lamp. However, all of the above mentioned methods present several important drawbacks; in particular, they are prone to operator bias, since image recording is not feasible, and do not permit the assessment of more sophisticated measurements of retinal vessels, such as the diameter of the vessels [2]. In addition, prior medical mydriasis is required, and presence of specially trained examiners is necessary. These important drawbacks have been counterbalanced by the development of advanced fundus cameras (mydriatic or non-mydriatic), allowing thus the accurate, objective, and repeatable representation of retinal blood vessels, and providing further insights in the diagnosis, classification and surveillance of retinopathy signs [3].

Detecting and measuring the spatial properties of the vessel network through automated techniques is a challenging task. From the perspective of the image processing, much effort has been devoted for segmenting the retinal vasculature and extracting its anatomical characteristics, such as width, length etc. A number of vessel detection and measurement techniques have been suggested in the literature (analytically described in [4]). In summary, the proposed approaches can be classified in two broad categories: i) methods that segment vessels from the image as a first step to estimate retinal structure, and ii) methods that track the vessel boundaries to extract their shape. The entire pool of the techniques that belong to the first category is generally further divided into two subgroups. In the first group, pixel-processing based techniques, such as local and global thresholding [5, 6], region growing [7] and matched filtering [8], fundamentally identify the vessel regions before extracting the available vasculature geometrical knowledge. More computationally intensive techniques based on 2D wavelet transforms [9], Gabor Wavelets mixed with classification [10], supervised learning [11-14] and active contours [15] may be also employed in this category. The second subgroup includes tracking methods [16, 17] that focus on the identification of the vascular network without the need to apply any segmentation step. Given an initial position, the vessel detection is achieved by tracking their boundaries. Fuzzy C-means clustering [18] has been effectively used since the extraction of the vasculature requires no prior assumptions about the characteristics and the edge information of the vessels.

The presented approach involves two separate modules. The first includes the detection and measurement of vessels in retinal images. The second module provides the ability to validate, edit and represent vessel information in multiple ways, by interactively selecting segments of interest and extracting their statistical information within spatial regions dictated by pertinent medical protocols. In brief, such methods include the automatic segmentation and measurement of a vessel-segment of interest, the computation of statistics across user-defined regions of medical interest, the interactive correction and editing of segmentation results,

ISBN: 978-606-544-078-4 413

Proceedings of the 3rd International Conference on E-Health and Bioengineering - EHB 2011, 24th-26th November, 2011, Iaşi, Romania___________________________________________________________________________________________________________________

Pre-Processing

Post-Processing

Image acquisition

High Resolution Low Resolution

Block Processing

Green Channel

Vessel Width Estimation

Optical Disc Detection

Pruning Small Vessel Branches

Statistical Results

Enhancement

Hessian-based Vessel Segmentation

Diffusion

Figure 1. Flowchart of the proposed framework.

the characterization of vessel segments with attributes (i.e. vain, artery) and other [19]. A Graphical User Interface (GUI) encapsulates the two modules and increases the automation of the measurement process in a user friendly manner.

The paper is structured as follows: Section II outlines the materials and methods used. Section III provides an overview of the experimental results on publically available datasets and section IV summarizes the key findings and provides general guidelines and pointers to future research.

II. MATERIAL AND METHODS

A. Experimental Data In order to verify the applicability of the proposed scheme,

our proposal was evaluated on two publicly available datasets of colored Red/Green/Blue (RGB) images and corresponding manual segmentations; the DRIVE [20], STARE [5] and an indicative dataset consisting of high resolution images, provided by the Hypertension Unit, of the 2nd Propaedeutic Department of Internal Medicine, of the Hippokration Hospital, of the Aristotle University of Thessaloniki, Greece. The DRIVE database consists of 40 retinal images along with manual segmentations of the vessels. The images were captured in digital form using a Canon CR5 non-mydriatic 3CCD. Their size is 768X584 pixels stored in JPEG format. These 40 images have been divided into training and test sets, containing 20 images each. The STARE database consists of 20 digitized slides captured by a TopCon TRV-50 fundus camera. The images were digitized to 700X605 pixels, in PPM file format. Finally, 10 high resolution images from the Hippokration Hospital were acquired with a NIDEK AFC-230/210 nonmydriatic digital fundus camera. The images acquired by this process are 2912X2912 pixel color images stored in JPEG format. B. Batch Image Processing

1. Pre-processing The applied techniques are presented in the following

paragraphs and illustrated as a pipeline of processes in Figure 1. Vessels appear more contrasted in the green channel; therefore the acquired retinal images are transformed from colored to monochromatic and the intensity of the green channel was used. Low resolution images were inverted before filtering so that vessels’ intensities are bright compared to the background.

Contrast limited adaptive histogram equalization (CLAHE) [21] is applied, to increase the global image contrast and reveal as much as possible the structural information of the vasculature. Finally, we incorporate a pre-processing step using an edge-preserving anisotropic diffusion filter [22] to smooth the images within homogeneous regions but without blurring vessel boundaries.

2. Segmentation Given the pre-processed image, the segmentation of blood

vessels is performed in two steps; vessel enhancement filtering and region-based identification of the vascular

network. A multiple scale filtering technique for vessel enhancement, based on the eigenvalue analysis of the Hessian matrix [23] is adopted in our study. The method starts by computing the second derivative of the image. As a second step, the gradient and the Hessian at multiple scales of the image are computed. Vessels with various sizes are then extracted based on the analysis of second order derivative features such as curvatures from the Hessian matrix. Using the eigenvalues of the Hessian matrix and the ratio between them, the vessel network is finally constructed.

After the vessel enhancement filter is completed, the contrast between vessels and other tissues is vastly improved. Finally, an iterative thresholding method for segmenting the blood vessel structure is then applied for the binarization of the enhanced image. A lot of effort has been made in order to provide an efficient and not-time consuming framework. Therefore, we focus on segmentation techniques that balance between accuracy and complexity and Otsu’s thresholding [24] was proven to act fairly well under those requirements.

3. Skeletonization A step, required for the characterization of the

morphological structure of the blood vessel's network, is the skeletonization of the segmented image, as in [25]. It generates one-pixel-wide skeletons by replacing the patterns that belong to vessels with line drawing representation of them. The skeletonized image isolated interesting features of the vasculature, such as a representation of the vessel “stem”, end-points, junction-points, and overlaps of vessels.

4. Block processing Segmenting high resolution images (2900X2900) can be a

computationally daunting task where most algorithms fail to achieve because of memory constraints. Instead of

414


Figure 2. Illustration of the methods used for the vascular network identification. The top left image is the original colored image. Top right is the extracted green channel of the original image. Middle left image shows the enhanced image after CLAHE and the middle right, the image after Hessian based vessel enhancement. In bottom, the left image is the image after segmentation and the right is the skeleton of the segmented image.

Figure 3. Pruning of spurious vessel branches. Top to bottom: original colored image with initial skeleton superimposed, and original colored image with pruned skeleton superimposed.

downsizing the available images a common way to overcome the high-dimensionality issues is image processing through distinct blocking. In distinct block processing, the image is divided into M-by-N blocks. These blocks, overlay the image matrix starting in the upper left corner, by adding some

border pads when the block does not fit exactly over it. Furthermore, the function pads overlapping borders to each block so that local operations such as spatial filtering can be seamlessly performed. Finally, each block is extracted from the image and passed to the segmentation algorithm. C. Post-processing

After the pre-processing steps and the segmentation of the images are completed (see Figure 2), a number of operations are automatically performed; the skeletonized images are processed to create an accurate representation of vessels. All the post-processing steps are further described below.

1. Size-based filtering The first stage of the post-processing procedure is to delete

very small isolated skeleton segments as they typically correspond to noise artifacts. The minimum component size is an input parameter, whose default value is determined proportionally to the resolution of the image; for the case of high resolution images it was set to 200 pixels. Segments below that value are eliminated. A typical 8-connected components process is applied to identify isolated elements in the skeleton image.

2. Pruning small vessel branches The segmentation results obtained from high resolution

images tend to exhibit richer structure at vessel boundaries than corresponding results from low resolution images. In particular a number of intrusions and protrusions can be observed which are typical due to combinations of image noise and low image contrast. In low resolution images, such artifacts have a smaller spatial extent and do not produce additional branches in the extracted skeleton structure. However, in high resolution images such structures give rise to spurious skeleton branches. The technique prunes “small” skeleton branches, as they correspond to such spurious branches, in order to facilitate more intuitive user interaction in the selection of branch segments (see Figure 3). Given as input the skeleton image and the maximum branch length mL as an input parameter, branches that have length smaller than mL are removed.

3. Vessel width estimation A local algorithm is employed to estimate the vessel width

at a skeleton point p. A hypothetical straight line segment T of length l is placed horizontally and centered upon p; length T is selected to be larger than any expected vessel width. Segment T is then consecutively rotated and in each rotation its intersections with the vessel's boundaries are located. The rotation that minimizes this length is assumed to bring T in such an orientation that intersects the vessel perpendicularly to its local orientation axis. The distance of the two intersections for this rotation is the measurement of vessel width at point p.

To find the intersections of T with the vessel for a given rotation, we initiate our query from point p. We then visit pixels in skeleton image in the direction of the rotated line segment, in both directions. The query, in each direction, ends when a non-vessel pixel (its value being 0) is found. When a user selects a point, the closest skeleton point to it, is

415


Figure 4. Top left and right image illustrates the vessel width measurement technique. Bottom image shows the width measurement of a vessel segment. The estimated boundaries for the indicated vessel segment (green dots) are displayed together with the skeletonization and the original colored image.

Figure 5. System overview. Top figure shows the vessel segments detected within a region of interest and a secondary characterization panel. Using this panel the user can characterize each segment, select the segments to be measured and export the analytical quantities to a file. Bottom figure illustrates the navigation utility panel of the user interface.

retrieved and the corresponding measurement, for that skeleton point, is provided. Vessel width is measured in pixels and can be converted to metric values if the image is calibrated with a reference object of known size. For this purpose, the application provides such an option and calculates the vessel widths in �m. Figure 4 illustrates the measurement technique of the vessels’ width.

4. Optical disc detection The optical disk is detected as the brightest, size dominant

blob in the image, as follows. Initially, several Mean Shift [26] processes are seeded equidistantly across the image. These processes use a circular kernel to converge to the local intensity maximum. Results leading to a low intensity value are disregarded. The remaining results are topologically clustered based on distance. The centroid C of the cluster with the largest cardinality is selected as the center of the optical disk. This process is inversely applied to detect the size-dominant darkest blob of a circular image [27].

To estimate the size of the disk the method proposed in [28] is adopted. The image neighborhood centered at C is convolved with a Laplacian of exponentially increasing sizes and the output is normalized by the scale factor (an exponential growth of 1.2 is utilized). The radius of the Laplacian at the scale where the maximum response is obtained is the estimate of the disks radius.

5. Statistical results The calculation of a statistical measure, introduced in [29],

has been integrated into our framework. In brief this measure calculates the ratio of two quantities, Central Retinal Arterial Equivalent (CRAE) and Central Retinal Venous Equivalent (CRVE), which are determined by the measurement of the arteries and veins enclosed in the region of interest respectively, according to the formulas developed in [30]. The mean widths of vein and artery segments within the region of interest are collected in two separate lists, namely the “Arteriole” and “Venule”. CRAE is computed as:

76.1022.001.1 2 −⋅⋅−⋅+⋅= bab2

a WWWW0.87CRAE , (1)

where Wb is the median value of “Arteriole” and Wa is the value in the same list exactly before the median. Correspondingly, CRVE is computed as:

05.45091.0 2 +⋅+⋅= b2

a WW0.72CRVE , (2) where now Wb is the median of “Venule” and Wa the value in the list exactly before Wb. Arteriovenus ratio (AVR) was also calculated as the ratio of these (CRAE/CRVE).

6. Integrated System The functionalities provided by the software are integrated

within the GUI. Beside the conventional logistic operations (i.e. select file or interest, export measurement results to file, etc.), the main goals of this GUI are: i) to support interactive measurements of medical professionals on regions and vessels of interest in an ergonomic way and to ii) facilitate

recovery from shortcomings of the automatic image processing algorithms and allow the expert user to modify their outcome.

The first goal is served by using an online representation of the detected vessels. In this way, the user can indicate vessels and vessel segments of interest with ease. At the same time, the GUI provides a visualization of this representation indicating to the medical professional, the detected vessels on

416


Figure 6. Boxplots illustrating the performance of the segmentation method when applied to DRIVE and STARE.

the original image. Using this visualization as feedback, the GUI supports the second goal by providing an editor where inaccurate vessel detection results can be corrected. The interactive measurement capabilities and the implementation of the GUI along with integration details are elaborated in [19]. Indicative system screens are shown in Figure 5.

III. SEGMENTATION RESULTS

To facilitate the comparison with other well-known retinal vessel segmentation approaches the performance of the proposed method was evaluated on the DRIVE and STARE datasets, using the accuracy, sensitivity and specificity of the vessel identification. These measures are proposed in the following equations.

FPTNTNyspecificit

FNTPTPysensitivit

FNFPTNTPTNTPaccuracy

+=

+=

++++=

(3)

True positive (TP) and true negative (TN) are the events when a pixel is correctly segmented as vessel or non-vessel respectively. False negative (FN) occurs when a vessel-pixel is segmented in the non-vessel area and false positive (FP) when a non-vessel pixel is segmented as a part of the vessel network. The test set from DRIVE was used for evaluation purposes. It had been manually segmented twice and the performance of our segmentation approach was measured using both segmentations as a gold standard. STARE dataset was also manually segmented by two observers and the segmentation performance was measured using both segmented image datasets.

Table I and Table II compare our approach results with some of the most popular algorithms for vessel using the two public datasets1. The performance of the proposed method was found to be similar and sometimes higher than the performance of more complex and time consuming techniques. The segmentation results are outlined as boxplots in Figure 6.

1 The manual delineation of the blood vessel network on the high

resolution images is an ongoing work, expected to be finalized in the coming months.

TABLE I. PERFORMANCE OF VESSEL SEGMETATION - DRIVE.

Method Sensitivity Specificity Accuracy Human Observer 0.7761 0.9725 0.9473

Supervised Methods Soares [10] 0.7230 0.9762 0.9446

Niemeijer [11] 0.6793 0.9801 0.9416 Staal [20] 0.7193 0.9773 0.9441

Unsupervised Methods Proposed 0.7414 0.9669 0.9371

Mendonca [4] 0.7344 0.9764 0.9452 Jiang [6] 0.6478 0.9625 0.9222 Perez [7] 0.7086 0.9496 0.9181

Chaudhuri [8] 0.2716 0.9794 0.8894 Vlachos [16] 0.7468 0.9551 0.9285

TABLE II. PERFORMANCE OF VESSEL SEGMETATION - STARE.

Method Sensitivity Specificity Accuracy Human Observer 0.8949 0.9390 0.9354

Supervised Methods Soares [10] 0.7103 0.9737 0.9480 Staal [20] 0.6970 0.9810 0.9516

Unsupervised Methods Proposed 0.7189 0.9656 0.9318

Mendonca [4] 0.6996 0.9730 0.9440 Perez [7] 0.7506 0.9569 0.9410

417


IV. CONCLUSION

In this study, an application that increases the automation of vessel measurement in fundoscopy images is presented. This application employs an image segmentation algorithm, along with pre-processing and skeletonization techniques in order to extract a representation of vessels. All the functionalities are integrated through a GUI, which assists the medical professional to perform measurements in an ergonomic fashion and apply targeted measurements according to medical protocols.

Future work will be pursued along the following directions. The first regards the improvement of the segmentation algorithm, in order to obtain even more accurate measurements. The second regards the temporal registration of fundoscopy images of the same patient, acquired across large time intervals (i.e. 6 to 12 months). This way, medical professionals will be able to automatically compare vessel changes, thus offering an objective tool towards a more efficient and accurate monitoring and diagnosis of hypertensive retinopathy.

ACKNOWLEDGMENT

This work was supported in part by the European Commission under the TUMOR (FP7-ICT-2009.5.4-247754) project.

REFERENCES [1] Grosso A., Veglio F., Porta M., Grignolo F., and Wong T. Hypertensive

retinopathy revisited: some answers, more questions. British Journal of Ophthalmology, 89(12):1646-1654, December 2005.

[2] Newsom R., Sullivan P., Rassam S., Jagoe R., and Kohner E. Retinal vessel measurement: comparison between observer and computer driven methods. Graefe's Archive for Clinical and Experimental Ophthalmology, 230(3):221-225, 1992.

[3] Chapman N., Witt N., Gao X., Bharath A., Stanton A., Thom S., and A. Hughes. Computer algorithms for the automated measurement of retinal arteriolar diameters. British Journal of Ophthalmology, 85(1):74-79, 2001.

[4] A. Mendonca and A. Campilho. Segmentation of retinal blood vessels by combining the detection of centerlines and morphological reconstruction. IEEE Transactions on Medical Imaging, 25(9):1200-1213, September 2006.

[5] A. Hoover, V. Kouznetsova, and M. Goldbaum, “Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response,” IEEE Trans. Med. Imag., vol. 19, no. 3, pp. 203–211, Mar. 2000.

[6] Xiaoyi Jiang; Mojon, D.; , "Adaptive local thresholding by verification-based multithreshold probing with application to vessel detection in retinal images," Pattern Analysis and Machine Intelligence, IEEE Transactions on , vol.25, no.1, pp.131-137, Jan. 2003.

[7] M.E. Martínez-Pérez, A.D. Hughes, A.V. Stanton, S.A. Thom, A.A. Bharath, and K.H. Parker, "Retinal Blood Vessel Segmentation by Means of Scale-Space Analysis and Region Growing", in Proc. MICCAI, 1999, pp.90-97.

[8] S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum,“Detection of blood vessels in retinal images using two-dimensional matched filters,” IEEE Trans. Med. Imag., pp. 263–269, 1989.

[9] J. J. G. Leandro, R. M. Cesar, Jr., and H. Jelinek, “Blood vessels segmentation in retina: Preliminary assessment of the mathematical morphology and of the wavelet transform techniques,” in Proc. 14th IEEE Comput. Soc. Brazil. Symp. Comput. Graph. Image Process. (SIBGRAPI), 2001, pp. 84–90.

[10] João V. B. Soares, Jorge J. G. Leandro, Roberto M. Cesar Jr., Herbert F. Jelinek, and Michael J. Cree, “Retinal Vessel Segmentation Using

the 2-D Gabor Wavelet and Supervised Classification,” IEEE Transactions on medical Imaging, vol. 25, no. 9, September 2006.

[11] Niemeijer M., Staal J., van Ginneken B, Loog M, and Abràmoff M. D., “Comparative study of retinal vessel segmentation methods on a new publicly available database,” in Proc. SPIE Med. Imag., M. Fitzpatrick and M. Sonka, Eds., 2004, vol. 5370, pp. 648–656.

[12] Ricci, E.; Perfetti, R.; , "Retinal Blood Vessel Segmentation Using Line Operators and Support Vector Classification," Medical Imaging, IEEE Transactions on , vol.26, no.10, pp.1357-1365, Oct. 2007.

[13] Marín, D.; Aquino, A.; Gegundez-Arias, M.E.; Bravo, J.M.; , "A New Supervised Method for Blood Vessel Segmentation in Retinal Images by Using Gray-Level and Moment Invariants-Based Features," Medical Imaging, IEEE Transactions on , vol.30, no.1, pp.146-158, Jan. 2011.

[14] Niemeijer, M.; Xu, X.; Dumitrescu, A.; Gupta, P.; van Ginneken, B.; Folk, J.; Abramoff, M.; , "Automated Measurement of the Arteriolar-To-Venular Width Ratio in Digital Color Fundus Photographs," Medical Imaging, IEEE Transactions on , vol.PP, no.99, pp.1,0.

[15] Al-Diri, B.; Hunter, A.; Steel, D.; , "An Active Contour Model for Segmenting and Measuring Retinal Vessels," Medical Imaging, IEEE Transactions on , vol.28, no.9, pp.1488-1497, Sept. 2009.

[16] M. Vlachos, E. Dermatas, Multi-scale retinal vessel segmentation using line tracking, Computerized Medical Imaging and Graphics, Volume 34, Issue 3, April 2010, Pages 213-227.

[17] Chutatape, O.; Liu Zheng; Krishnan, S.M.; , "Retinal blood vessel detection and tracking by matched Gaussian and Kalman filters," Engineering in Medicine and Biology Society, 1998. Proceedings of the 20th Annual International Conference of the IEEE, vol.6, no., pp.3144-3149 vol.6, 29 Oct-1 Nov 1998.

[18] Tolias, Y.A.; Panas, S.M.; , "A fuzzy vessel tracking algorithm for retinal images based on fuzzy clustering," Medical Imaging, IEEE Transactions on , vol.17, no.2, pp.263-273, April 1998.

[19] P. Karamaounas, G. Manikis and X. Zabulis.Retinal Images Analyzer. Institute of Computer Science-FORTH: Heraklion Greece. 2011:TR416

[20] Staal, J.; Abramoff, M.D.; Niemeijer, M.; Viergever, M.A.; van Ginneken, B.; , "Ridge-based vessel segmentation in color images of the retina," Medical Imaging, IEEE Transactions on , vol.23, no.4, pp.501-509, April 2004.

[21] K. Zuiderveld. Contrast limited adaptive histogram equalization. Graphics gems IV, pages 474-485, 1994

[22] P. Perona, J. Malik, “Scale-space and edge detection using anisotropic diffusion,” IEEE Trans. PAMI, V. 12, no. 7, pp. 629-639, July 1990.

[23] A.F. Frangi, W.J. Niessen, K.L. Vincken, M.A. Viergever (1998). Multiscale vessel enhancement filtering. In Medical Image Computing and Computer-Assisted Intervention - MICCAI'98, W.M. Wells, A. Colchester and S.L. Delp (Eds.), Lecture Notes in Computer Science, vol. 1496 - Springer Verlag, Berlin, Germany, pp. 130-137.

[24] Otsu, N., "A Threshold Selection Method from Gray-Level Histograms," IEEE Transactions on Systems, Man, and Cybernetics, Vol. 9, No. 1, 1979, pp. 62-66.

[25] Lam L., S. Lee, and C. Suen. Thinning methodologies-a comprehensive survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 14(9):879, September 1992.

[26] Y. Cheng. Mean shift, mode seeking, and clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 17(8):790-799, 1995.

[27] X. Zabulis, A. A. Argyros, and D. P. Tsakiris. Lumen detection for capsule endoscopy. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3921-3926, 2008.

[28] T. Lindeberg. Detecting salient blob-like image structures and their scales with a scale-space primal sketch: A method for focus-of-attention. International Journal of Computer Vision, 11:283-318, 1993.

[29] J. Parr and G. Spears. Mathematic relationships between the width of a retinal artery and the widths of its branches. American Journal of Ophthalmology, 77:472-477, 1974.

[30] Hubbard LD, Brothers RJ, King WN, Clegg LX, Klein R, Cooper LS, et al. Methods for evaluation of retinal microvascular abnormalities associated with hypertension/sclerosis in the Atherosclerosis Risk in Communities Study. Ophthalmology;106(12):2269-80, Dec 1999.

418


Proceedings of the 3rd International Conference on E ...zabulis/EHB11_Retina_Analysis.pdf · 2. Segmentation Given the pre-processed image, the segmentation of blood vessels is performed

Documents