Segmentation of optic disc and vessel based optic cup for glaucoma Assessment

Transcription

1 Segmentation of optic disc and vessel based optic cup for glaucoma Assessment 1 K.Muthusamy, 2 J.Preethi 12 (CSE, Anna University Regional Center Coimbatore, India, 1 muthuksp105@gmail.com, 2 preethi17j@yahoo.com) Abstract Glaucoma is a chronic eye disease that leads to vision loss. In this disease, the optic nerve is progressively damaged. Detection of this Glaucoma is very difficult task and Current tests using intraocular pressure (IOP) are not sensitive enough for population based glaucoma screening. The optic nerve head is also called the optic disc and central bright zone called the optic cup, Optic nerve head assessment in retinal images is more difficult. This paper proposes classification of optic disc based on super pixel and optic cup using vessel bends tracking for glaucoma screening, In optic disc segmentation, histograms is applied to R,G,B,H and S color space, and center surround statistic is calculated to classify each superpixel as disc or background. A self assessment reliability test is performed to evaluate the quality of the automated optic disc segmentation. In optic cup vessel bends tracking is also included to fine tune the optic cup boundary. Vessel bends are identified using the method of dynamic Region of Support (ROS), optic cup is segmented and also location information is added. Finally segmented optic disc and cup is used to calculate Vertical Cup to Disc Ratio (CDR). CDR is one of the glaucoma factors and CDR is well accepted and commonly used. If CDR value is high then risk of glaucoma. This method can be useful for automatic segmentation and glaucoma screening. Keywords Glaucoma screening, optic disc segmentation, optic cup segmentation, vessel bends, Region Of Support. 1. INTRODUCTION GLAUCOMA is a major eye disease in the world. in this disease optic nerve is damaged. glaucoma is the second leading cause of blindness, and it is determined to affect more than 80 million people by Increasing of the disease leads to vision loss, which occurs over a long period of time. when the disease is quite advanced the symptoms will occurs, and it is also called as silent thief of sight. Glaucoma cannot be cured, but progression of this glaucoma can be slowed down by various treatment. Therefore, detecting glaucoma in time is critical. However, several glaucoma patients are not known of the disease until it has reached its final stage. In Singapore, more than 95% of patients are unaware that they have this condition. more than 70% of people with this disease are undiagnosed. Since glaucoma progresses with few signs or symptoms and the loss of vision from glaucoma disease is irreversible. There are 3 mechanisms to detect glaucoma: 1) assessment of raised intra ocular pressure (IOP), 2) assessment of damaged optic nerve head. 3) assessment of abnormal visual field, The IOP measurement using non contact tonometry (also known as the airpuff test ) is neither specific nor sensitive enough to be an effective screening tool because glaucoma disease can be present with or without increased IOP. A functional test requires special equipments only present in territory hospitals and therefore unsuitable for screening. Assessment of the damaged optic nerve head is both more difficult, and superior to IOP measurement or visual field testing for glaucoma screening. Optic nerve head assessment can be done only by a trained professional. However, manual assessment is sensitive, time consuming and more expensive. Therefore, automatic optic nerve head assessment will be very useful. One method for automatic optic nerve head detection is to use image features for a binary classification between glaucomatous and healthy one. These features are normally obtained at the level of image. In these strategy, features selection and classification method is difficult and challenging [8]. The other method is to follow the clinical glaucoma indicators. Main glaucoma risk factors are considered, such as the vertical cup to disc ratio (CDR), disc diameter, ISNT [11] rule, peripapillary atrophy (PPA), notching etc. Although several ophthalmologists have different opinions on the usefulness of these factors, CDR is well and most accepted and commonly used method. A larger CDR indicates a higher level risk of glaucoma. There are many research of automatic CDR measurement from fundus images. because 3-D fundus images are not available immediately, 2-D color fundus images are still used to by most clinicians [8],.This paper focuses on automatic segmentation of optic disk and cup for glaucoma screening using CDR from 2-D fundus images. The optic nerve head is also called the optic disc is located where ganglion cell axons exit the eye to form the optic nerve, through which visual information is transmitted to the brain. In 2-D fundus images, the disc can be divided into two parts; one is central bright zone called the optic cup (in short, cup) and a one is peripheral region called the neuroretinal rim. Fig. 1 shows the main structures of the disc. Calculation of CDR is Volume: 01 Issue: 05 May

2 based on the ratio of the vertical cup diameter (VCD) to vertical disc diameter (VDD). Fig 1. Structure of optic disc, optic cup and CDR calculation. to calculate CDR good segmentations of disc and cup are essential for CDR measurement. There are many strategy [8], have been proposed for automatic CDR measurement from 2- D fundus images. This paper proposes segmentations of optic disc based on superpixel classification and segmentation of optic cup based on vessel bend tracking for glaucoma screening. A same concept has been used for vessel segmentation [2]. Using superpixels, compute center surround statistics and unify them using histograms equalization for disc and cup segmentation. it can get prior information of the cup by computing location information for cup segmentation. CDR is computed based on the segmented disc and cup for glaucoma screening. And also, the proposed model computes a self-assessment reliability score for its accuracy of disc segmentation result. Self-assessment is more important and main issue of disc segmentation. In practical, an automatic segmentation method might work well for most images while working poorly for the rest. Therefore, it is very much important to have self-assessment where users are warned of cases with potentially occurrence of large errors. This paper is organized as follows. In Section II, perform optic disc segmentation based on the superpixel classification and also superpixels generation is included in this section, and features extraction operation is performed from superpixels for the classification and the self-assessment reliability score is computed. In Section III perform classification of cup segmentation on based superpixel. this procedure is same to disc segmentation. In Section IV shows the CDR calculation for glaucoma screening is performed. Finally in section V discussions are made about this paper advantage, conclusions, future works and reults are presented. II. SEGMENTATION OF OPTIC DISC Segmentation of disc is so important in many computer aided diagnosis systems, also in glaucoma screening. The localization is used to find disc pixel, very often the center. This paper is to solve segmentation problem. The segmentation process determine the disc boundary, which is very difficult job due to occlusions blood vessel, pathological changes across disc area, variable imaging conditions, etc There are some approaches have been proposed for disc segmentation, which are classified as template based methods, deformable model based methods and pixel classification based methods [14]. Disc boundary is modeled using circular Hough transform. Because of its computational efficiency. However, clinical studies show that a disc has a slightly oval shape with the vertical diameter being about 7% 10% larger than the horizontal one. Active contour model is used to find the optimal points based on the image gradient and the smoothness of the contour. in deformable model technique is used to minimize the energy function by image gradient, image intensity and boundary smoothness. A level set is used to estimate the disc followed by an ellipse fitting to smooth the boundary. To extract the disc boundary two methods edge detection and circular Hough transform are combined with an active shape model. the active shape model is also applicable for on probability maps to find the disc boundary. In addition, we also adopt a classification of superpixel based approach using histograms. It is very useful to initialization of the disc for deformable models. The PPA region will be confused as part of disc For two reasons 1) PPA region looks similar to the disc; 2) its shape makes it form another ellipse together with the disc (often stronger). Pixel classification based methods and deformable model based methods limitations are solved by using superpixel classification based method and this method is combined with the deformable model based methods. and this method reduce the complexity of subsequent processing by capture redundancy in the image. In this paper, superpixel classification is used for an initialization of disc boundary disc boundary is determined by using the deformable model. A. Generation of Superpixel There are several algorithms proposed for superpixel generation for many images such as various images of scene, animal, human, etc. this paper use the simple linear iterative clustering [3] algorithm (SLIC) to generate superpixels by aggregating nearby pixels into superpixels in images. SLIC algorithm is fast one, memory efficient and has great boundary adherence. SLIC is also simple to use [3]. In SLIC, initial cluster centers are sampled on a regular grid spaced by S= pixels. The centers are moved towards the lowest gradient position in a 3 3 neighborhood. Clustering is then applied. For each SLIC iteratively searches for its best matching pixel from the 2S neighborhood around based on color and spatial proximity and then compute the new cluster center based on the found pixel. The iteration continues until the distance Volume: 01 Issue: 05 May

3 between the new centers and previous ones is small enough. Finally, a post processing is applied to enforce connectivity. Fig 2. B. Extraction of Feature Overall process 1) Contrast Enhanced Histogram: There are several features such as color, appearance; gist, location and texture can be extracted from superpixels for classification. Detection of main differences between disc and non-disc region using color histogram from superpixels is a great choice. Histogram equalization operation is applied to red, green, and blue channels from RGB color spaces individually to enhance the contrast. Histogram equalization may leads to dramatic changes in the image s color balance. Thus, hue and saturation from HSV color space are also added to form five channel maps. The histogram of each superpixel computed from all the five channels: the histogram equalized r, g, and b as well as the original h,s. The histogram computation uses 256 bins and 256 5=1280 dimensional feature HIST j =[ ] (1) is computed for the jth super pixel, where HE(.) denotes the function of histogram equalization and the function to compute histogram from 2) Center Surround Statistics: The PPA region will be very close to the disc. It is very important to add features that shows the difference between the PPA region and the disc region. The superpixels from the two regions often appear similar except for the texture: the PPA region contains blob-like structures while the disc region is relatively more homogeneous. The superpixel histogram does not work well as the texture variation in the PPA region is often from a larger area than the superpixel. Superpixel consists of a group of pixels with similar colors. So this paper use center surround statistics (CSS) from superpixels as a texture feature. To compute CSS, nine spatial scale dyadic Gaussian pyramids are generated with a ratio from 1:1 (level 0) to 1:256 (level 8), Multiple scales are used as the scale of the blob-like structures largely vary The dyadic Gaussian pyramid is a hierarchy of low-pass filtered versions of an image channel. It is accomplished by convolution with a linearly separable Gaussian filter and decimation by a factor of two. Then center surround operation between center (finer) levels c = 2,3,4 and surround levels (coarser) S = c + d with d = 3,4 is applied to obtain six maps empirically computed at levels of 2 5, 2 6, 3 6, 3 7, 4 7, and 4 8 from an image channel. Denote the feature map in center level c as I(c) and the feature map in surround level s as I(s). first interpolate I(s) to be the same size as I(c ) and the interpolated map is denoted as (I(s)), where (.) denotes the interpolation from the surround level s to the center level c. The center surround difference is then computed as. All the difference maps are resized to be the same size as the original. Compute the maps from and channels to get 6x 3=18 maps. C. Classification of Optic disc This paper use Support Vector Machine (SVM) as a classifier. first create active training data set to be a subset of the available training data set (pool) and all training iteration is made on the the active set. This returns a preliminary classifier. Then classifier is used as a pool evaluator. training set is increased in every round of training by misclassified in the previous iteration and continue the process until there is no new action in the classification accuracy or the maximum iterations have been reached. Here it is not directly using the binary classification results from SVM, decision function are (2) (3) Volume: 01 Issue: 05 May

4 used from SVM output values. Each superpixel output is used as the decision values for all pixels. After that smoothed decision value is obtained by mean filter. Then a binary decision for all pixels with a threshold is obtained using smoothed decision values. Assign +1 and -1 to positive (disc) and negative (non-disc) samples, and the threshold is the average of them. D. Self-Assessment Reliability Score if the superpixel based segmentation is very close to the actual boundary then the obtained boundary values from classifier before and after ellipse fitting must be close. Else, the result of above is to be less reliable. So compute Self Assessment of optic disc segmentation. Define the set of raw estimation points as X and the set of fitted estimation points as Y = f(x). and obtain its nearest point in Y and their distance is computed as (4) III. SEGMENTATION OF OPTIC CUP One of main challenging task is to detecting the cup boundary from 2-D fundus images without depth information. Pallor is one of landmark to determine region of cup. And another landmark is blood vessel bends inside the cup boundary. Detection these vessel bends are very important but it is not a easy task in cup. In this paper vessel bends tracking method is used to detect the cup boundary. A. Feature Extraction The same feature that was extracted in disc segmentation is used here that is histogram equalization is applied to r, g and b color channels and also h, s color channels added to form a five color channels and centre surround statistics (CSS) is calculated as in disc segmentation and finally this paper include location information that is calculated as following. The distance Dj between center of the jth superpixel and center of the disc is called as location information. Mathematically Dj is calculated as Where denotes the coordinate of the disc center., denotes the coordinate of the center of SPj and denotes the height and width of the disc, respectively. B. Segmentation of Cup Using r-bends Information This paper focusing only on segmentation of cup using vessel bends. The vessels occurring in several points through optic disc and subset of this vessel are present inside the cup boundary. There are two type of vessels one is thin and thick vessels. This vessel bends affecting the cup boundary. So detection of these vessel bends present in the cup boundary is (5) very difficult task. This relevant vessel bends are known as r- bends. 1) Medial Axis Detection: There are so many methods to detect the vessel bends in cup boundary, in this paper method proposed in [48], which formulates the blood vessel detection as a problem of trench detection in the intensity surface.the final trench points gives medial Axis Detection representation of vessel structure. 2) Vessel Bend Detection: In which,the amount of vessel bends is based on the type of vessels. Thin vessel having significant bends compared to thick vessels. To detect these vessel bends this paper employ the method based on dynamic Region of Support (ROS). First, extract vessel segments terminated by end and/or junction points. For each segment and compute 1- D shape (curvature) profile and locate the local maxima. A ROS for any is defined as a segment of vessel around and bound on either side by the nearest curvature minimum. A sector is radially analyzed with a step size of 20 and in each step, only bends formed by vessels with the correct orientation are retained. If multiple bends remain, then the bend with smaller value of is selected as thin, rather than thick. Vessel bends are more reliable indicators for the cup boundary. IV. CALCULATION OF CUP TO DISC RATIO There are many number of glaucoma factors to screening the disease. Cub to disc ratio is one of main glaucoma factors. CDR is mostly acceptance and used factors to screening the glaucoma. After segmented both optic disc and optic cup, CDR is calculated based on the Vertical Disc Diameter VDD and Vertical Cub Diameter VCD as follows CDR= (6) Computed CDR value is greater than the threshold value then its indicating high risk of glaucoma. V. DISCUSSIONS AND CONCLUSION In this paper segment the optic disc based on the superpixel classification, in which histogram equalization is applied to R, G, B and H, S color channels also included to form five channels to enhance contrast of fundus image. Then center surround statistics (CSS) is calculated to fine tune the disc boundary finally self assessment reliability score is computed to verify the correctness of automated segmentation of optic disc and then optic cup is segmented based on the vessel bends present inside the cup boundary using the method of Volume: 01 Issue: 05 May

5 dynamic Region of Support (ROS). In which location information Dj is also calculated to find the center of cup boundary and finally CDR is calculated to indicates the level of glaucoma. In future, glaucoma factors can include to screening the glaucoma disease efficiently and another future work is using multiple kernels to classify the optic disc and optic cup effectively. RESULTS Fig 5. R,G,B histogram Equalization Fig 3. Before superpixel generation FIG 6 VESSEL TREE SEGMENTATION Fig 4. After Superpixel Generation FIG 6.OPIC CUP SEGMENTATION USING VESSEL BENDS Volume: 01 Issue: 05 May

6 REFERENCES [1] Jun Cheng*, Jiang Liu, Superpixel Classification Based Optic Disc and Optic Cup Segmentation for Glaucoma Screening IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 32, NO. 6, JUNE [2] G. D. Joshi, J. Sivaswamy, and S. R. Krishnadas, Optic disk and cup segmentation from monocular color retinal images for glaucoma assessment, IEEETrans. Med. Imag., vol. 30, no. 6, pp , Jun [3] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. Susstrunk, Slic superpixels compared to state-of-the-art superpixel methods, IEEE Trans. Pattern Anal. Mach. Intell., vol. 34, no. 11, pp , Nov [4] W. K. Wong, J. Liu, J. H. Lim, X. Jia, H. Li, F. Yin, and T. Y. Wong, Level-set based automatic cup-to-disc ratio determination using retinal fundus images in argali, Proc. Int. Conf. IEEE Eng. Med. Biol. Soc., pp , [5] M. Foracchia, E. Grisan, and A. Ruggeri, Detection of optic disc in retinal images by means of a geometrical model of vessel structure, IEEE Trans. Med. Imag., vol. 23, no. 10, pp , Oct [6] Aquino, M. Gegundez-Arias, and D. Marin, Detecting the optic disc boundary in digital fundus images using morphological, edge detection, and feature extraction techniques, IEEE Trans. Med. Imag., vol. 29, no. 11, pp , Nov [7] Z. Zhang, J. Liu, N. S. Cherian, Y. Sun, J. H. Lim, W. K.Wong, N. M. Tan, S. Lu, H. Li, and T. Y. Wong, Convex hull based neuro-retinal optic cup ellipse optimization in glaucoma diagnosis, in Int. Conf. IEEE Eng. Med. Biol. Soc., 2009, pp [8] O. Veksler, Y. Boykov, and P. Mehrani, Superpixels and supervoxels in an energy optimization framework, in Proc. Eur. Conf. Comput.Vis., 2010, vol. 5, pp [9] G. D. Joshi, J. Sivaswamy, and S. R. Krishnadas, Optic disk and cup segmentation from monocular color retinal images for glaucoma assessment, IEEE Trans. Med. Imag., vol. 30, no. 6, pp , Jun [10] L. Itti, C. Koch, and E. Niebur, A model of saliencybased visual attention for rapid scene analysis, IEEE Trans. Pattern Anal. Mach. Intell., vol. 20, no. 11, pp , Nov [11] J. Cheng, D. Tao, J. Liu, D. W. K. Wong, N. M. Tan, T. Y. Wong, and S. M. Saw, Peripapillary atrophy detection by sparse biologically inspired feature manifold, IEEE Trans. Med. Imag., vol. 31, no. 12,pp , Dec [12] D. Joshi, J. Sivaswamy, K. Karan, and R. Krishnadas, Optic disk and cup boundary detection using regional information, in Proc. IEEE Int. Symp. Biomed. Imag., 2010, pp [13].N. Harizman, C. Oliveira, A. Chiang, C. Tello, M. Marmor, R. Ritch, and J. M. Liebmann, The ISNT rule and differentiation of normal from glaucomatous eyes, Arch. Ophthalmol., vol. 124, pp , 2006 [14] N. Inoue, K. Yanashima, K. Magatani, and T. Kurihara, Development of a simple diagnostic method for the glaucoma using ocular fundus pictures, in Int. Conf. IEEE Eng. Med. Biol. Soc., 2005, pp [15] Z. Zhang, F. Yin, J. Liu, W. K. Wong, N. M. Tan, B. H. Lee, J. Cheng, nand T. Y. Wong, Origa-light:An online retinal fundus image database for glaucoma analysis and research, in Int. Conf. IEEE Eng. Med. Biol. Soc., 2010, pp Volume: 01 Issue: 05 May