CHAPTER 4 OPTIC CUP SEGMENTATION AND QUANTIFICATION OF NEURORETINAL RIM AREA TO DETECT GLAUCOMA

Transcription

1 83 CHAPTER 4 OPTIC CUP SEGMENTATION AND QUANTIFICATION OF NEURORETINAL RIM AREA TO DETECT GLAUCOMA 4.1 INTRODUCTION Glaucoma damages the optic nerve cells that transmit visual information to the brain. Though IOP is the most significant reason to develop glaucoma, optic disc parameters are mainly used to discover early stage of glaucoma damage. So there is a real need for CAD methods that can identify early glaucomatous development to avoid further progression. In this chapter, a technique for cup segmentation to detect glaucoma using Color Mathematical Morphology (CMM) is proposed. Structural features are then described to estimate quantitative parameters to classify fundus images into two classes. Cup to disc ratio and the rim area in Inferior, Superior, Nasal, Temporal (ISNT) quadrants are the parameters extracted from the segmented disc and cup region. Flow diagram of the proposed technique to detect glaucoma using structural features are provided in section 4.2.Section 4.3 describes an automatic optic cup segmentation algorithm using color space and mathematical morphology. Sections 4.4 and 4.5 describe the experimental results and performance analysis for optic cup. Parameter estimation for glaucoma assessment, neuroretinal rim quantification and rim area detection algorithm are presented in section 4.6. Performance analysis of structural features is explained in section 4.7. Conclusions are provided in section 4.8.

2 ASSESSMENT OF GLAUCOMA USING STRUCTURAL FEATURES Optic cup is segmented using CMM technique based on pallor in fundus images to differentiate the optic cup from the disc boundary. From the detected disc and cup boundary, parameters namely cup to disc ratio and neuroretinal rim area are calculated. Flow diagram to detect glaucoma using the structural features is shown in Figure 4.1. Fundus image Preprocessing Optic disc segmentation Optic cup segmentation Segmented features (CDR, Neuroretinal rim area) (CDR, Neuroretinal Identification of stages of of glaucoma glaucoma Figure 4.1 Flow diagram to assess glaucoma using structural features 4.3 PROPOSED CMM METHOD TO DETECT OPTIC CUP Optic cup detection is a challenging task because of the interweavement of cup with the blood vessels. Optic disc is detected using

3 85 region of interest based segmentation and the bounding rectangle enclosing the region of interest and it is set as 1.5 times the disc width parameter. Color mathematical morphology approach for the segmentation of optic cup is aimed to detect the optic cup exactly to calculate the neuroretinal rim area present between the disc and cup. Unlike most of the previous methods discussed in the literature, the method proposed here differs in the initial detection of optic cup followed by the erasure of blood vessels to get higher accuracy. A new color model technique based on pallor in fundus images using K means clustering as shown in Figure 4.2 is proposed to differentiate the optic cup from the disc boundary. Figure 4.2 Systematic representation of CMM model to detect optic cup Optic cup is detected using the technique proposed in Figure 4.3.

4 86 Input Image Optic Disc Masking K-Means Clustering Centroid Color Mapping Cup Color Identifying Cup Boundary Extraction Morphological Operations Ellipse Fitting Figure 4.3 Flow diagram for optic cup detection The optic cup and disc areas usually differ in color, known as pallor. This method makes use of this difference in pallor to delineate the cup-disc boundary. Observations on the retinal images show that the actual cup pallor differs between different patients and even between images of the same retina due to changes in the lighting conditions. So color intensity of the optic cup cannot be fixed with prior knowledge. To overcome the limitation of the RGB color space in high level processing, color spaces have been developed based on the mathematical transformation of the original RGB color channels.

5 87 In a color image, each pixel has a 3D vector with component values ranging from 0 to 255. RGB, YIQ, CMY, XYZ, HSV, LUV and L* a* b* are the color models oriented towards image processing application. Color fundus images of Red, Green and Blue (RGB) space are transformed into Commission International d Eclairage (CIE) Lab color space, which is a color-opponent space with dimension L for lightness, a and b for the coloropponent dimensions based on nonlinearly-compressed CIE XYZ color space coordinates. Euclidean distance in L* a* b* is the threshold for human to distinguish color difference. CIE L* a* b* color space is perceptually uniform and is device independent. It consists of a luminosity layer L* which varies uniformly from 0 for black to 100 for white, chromaticity-layer a* indicating where color falls along the red-green axis, and chromaticity-layer b* indicating where the color falls along the blue-yellow axis. The a and b values are expressed such that +a/-a denotes red/green and +b/- b denotes blue/yellow. This color space is an excellent decoupler of intensity and color. Non-linear relations described by (William K.Pratt 2002) for L* a* and b* are as follows: L = 116 f 16, if > (4.1) a = 500 f f (4.2) b = 200 f f (4.3) where f (t) = t if t > t + otherwise The division of the f(t) function into two domains was done to prevent an infinite slope at t = 0.Tristimulus values represent the relative quantities of the primary colors. The intensities of red, green and blue are

6 88 transformed into tristimulus values and are represented by X, Y and Z. Here X n, Y n and Z n define the illuminant dependent reference white point in the XYZ color space where the subscript n' denotes normalized values. In order to detect the optic cup region from the surrounding region, a segmentation algorithm based on color histogram analysis followed by morphological operations has been developed. Proposed CMM color model for the detection of optic cup is explained in the following steps. 1. Input image to the optic cup segmentation process is the segmented optic disc in L *a* b* color space. 2. Optic disc is masked with a radius equal to or greater than the size of the optic disc. 3. The masked image is fed to the clustering process in order to group the pixel values into regions.number of clusters for the clustering process is determined using anatomical knowledge as well as Hill Climbing technique. 3.1 As optic disc consists of four regions viz optic cup, interior optic disc, exterior optic disc and blood vessels, the number of clusters is selected as four (K=4) using domain knowledge. 3.2 Number of clusters for K means clustering is also determined automatically using Hill Climbing technique described by Ohashi et al (2003). The clusters are found using the peaks by projecting the image onto its three color components. Histograms are constructed by splitting the range of data into equal sized bins called classes. Then for each bin, the number of points from the data set that falls into each bin are counted. Since CIE L*a*b* feature space is three dimensional, each bin in the color histogram has B d -1 neighbours where B is the total

7 89 number of bins, chosen as 10 by trial and error method and d, the number of dimensions of the feature space is 3. Peaks are identified by comparing the pixels with the neighboring bins and the number of peaks obtained indicates the value of K. The histogram bins are utilized rather than the pixels themselves finding the peaks of clusters D color histogram of the image is computed A non-zero bin of the color histogram is started and an uphill move is made until reaching a peak as follows: i. The number of pixels of the current histogram bin is compared with the number of pixels of neighboring left and right bins. ii. iii If the neighboring bins have different numbers of pixels, the algorithm makes an uphill move towards the neighboring bin with larger number of pixels. If the immediate neighboring bins have the same numbers of pixels, the algorithm checks the next neighboring bins, and so on, until two neighboring bins with different numbers of pixels are found. Then, an uphill move is made towards the bin with larger number of pixels The uphill climbing is continued until reaching a bin from where there is no possible uphill movement. Hence, the current bin is identified as a peak (local maximum). This step is continued until all non-zero bins of the color histogram associated with a peak are climbed. The identified peaks as shown in Figure 4.4

8 90 represent the initial number of clusters of the input image and these peaks are saved. Number of clusters K is initialized as four (K=4), and the values of these bins form the initial seeds as shown in Figure 4.4. These formed seeds are then passed on to K means clustering. Lower value of K leads to an increase in the cup size and higher value results in the predominance of blood vessels. Figure 4.4 Detected peaks in L*a*b* color space x 1,x 2,...x M are the data points or vectors of observations. Each vector x i will be assigned to one and only one cluster. M data points are grouped into K clusters such that similar items are grouped together in the same cluster. For each data point, nearest centroid is found and the data point

9 91 is assigned to the cluster associated with the nearest centroid. The new centroid will be the mean of all points in the cluster. Centroid is taken and data is mapped to the closest one, using the absolute distance between them. The above two steps are iterated until convergence and when there are no new re-assignments it is stopped. For a current set of cluster means, each observation is assigned as in Equation (4.4). C(i) = arg min x m, i = 1,, M (4.4) C(i) denotes the cluster number for the i th observation, M denotes the number of data points in the cluster, number of clusters K is initialized as 4, m k is the mean vector of the k th cluster., where k varies from 1 to 4. K-means minimizes within-cluster point scatter as shown in Equation (4.5). (4.5) 4. By using K means clustering all the peaks of three channels are grouped into one cluster and valleys, pixels into separate clusters. K- Means clustering groups the pixels within the optic disc into the four regions. 5. Each cluster has a centroid. Then each region is filled with the corresponding region s centroid color. From these 4 regions, the region corresponding to optic cup can be easily identified by its centroid color. Each pixel within a cluster is then replaced by the corresponding cluster centre color. 6. Difference between two colors are measured using Euclidean distance metric. The colors {c 1,c 2,c 3, c 4 } are extracted from the image. The first

10 92 cluster centre a 1 in the color space is chosen as a 1 =c 1.Next the color difference from c 2 to a 1 is computed. If this difference is greater than a threshold T h, a new cluster centre is created as a 2 =c 2, otherwise c 2 is assigned to a 1. Similarly the color difference from each representative color to every established color centre is computed and thresholded. A new cluster is created if all of these distances exceed T h, otherwise color is assigned to the class to which it is closest. The brightest centroid color corresponds to an optic cup. Thus an initial boundary of an optic cup is obtained and the output of this clustering is an optic cup. Experiment is repeated for various bin values like 5, 10 and 15 as shown in Figure 4.5. (a) Input Image (b) Clustered outputs for N =5 (c) Clustered outputs for N=10 (d) Clustered outputs for N=15 Figure 4.5 Clustered outputs for various bin size

11 93 The impact of blood vessel region within the cup is removed by morphological operations. This is performed by a closing operation, that is, a dilation to first remove the blood vessels followed by erosion operation to restore the boundaries to their former positions in the region of interest. As gray level morphological operations cannot be directly applied to color images, each color retinal image is described as a set of three independent vectors and each of these vectors represent a gray scale image. In color morphology each pixel must be considered as a vector of color components. Definitions of maximum and minimum operations on ordered vectors are necessary to perform basic operations. Hence for each arbitrary point x in the color space, morphological dilation (I d ) and erosion (I e ) function by a structuring element E are defined as in Equations (4.6) and (4.7). Dilation of the image I by B I (x) = {I(y): I(y) = max [I(z)], z E (4.6) Erosion is defined by I (x) = {I(y): I(y) = min[i(z)], z E (4.7) Structuring element is a matrix of pixels, each with a value of zero or one. The pixel in the structuring element containing 1 s defines the neighborhood of a pixel. When the structuring element fits the image, for each of its pixels set to 1, the corresponding image pixel is also 1. Similarly, a structuring element is said to hit an image if, at least for one of its pixels set to 1 the corresponding image pixel is also 1. The closing operator usually smoothes away the small scale dark structures from the color retinal images. As closing eliminates the image details smaller than the structuring element, the structuring element E must be set to cover all possible vascular structures. A circular window of maximal vessel width as radius is used for dilation and

12 94 erosion. Of the various shapes and sizes like square, rectangular, diamond, diagonal shaped, a disc shaped structuring element of size is preferred since blood vessels were determined not to be wider than 13 pixels. Size of the structuring element depends on the features extracted from the image. A symmetrical 13x13 disc-structuring element is used, since the blood vessels were determined to be not wider than 13 pixels. In the output image, pixel computation is performed by taking the maximum of the input pixels for dilation and minimum of the pixels in the neighborhood for erosion.this step helps to reject outliers inside or outside the cup region and helps to get approximate cup region. For images having a large traversal of blood vessels and discontinuous cup region, ellipse fitting technique is used. Ellipse fitting based on least squares fitting algorithm is used to smooth the cup boundary. The modified optic cup boundary obtained is then fitted with ellipse as shown in Figure 4.6. (a) Input image (b) Mask image (c) Color model (d) Initial cup boundary (e) Image smoothing (f) Ellipse fitting Figure 4.6 Steps in the detection of optic cup

13 95 Further, the boundary of the cup is detected exactly in accordance with the markings of experts except in those regions where the optic cup boundary is severely occluded by the blood vessels. A comparison of color spaces for cup detection is shown in Figure 4.7. Rand Index (RI), a statistical measure is used to measure the quality of segmentation. RI counts the fraction of pixels whose labeling are consistent between the computed segmentation and ground truth. RGB, YIQ, HSV, XYZ color space exhibits a RI of 0.883, 0.796, and respectively. In these color spaces, segmented cup results do not show consistent results with the ground truth and hence the RI is lower. In L*a*b*color space, optic cup is well segmented and exhibits a high RI of 0.93 when compared with other color spaces. Figure 4.7 Optic cup detection in different color spaces

14 EXPERIMENTAL RESULTS A few input images are shown in Figure 4.8. Figure 4.9 illustrates the detected disc, cup boundary and Figure 4.10 shows the neuroretinal rim region present between the disc and the cup. Figure 4.8 Few input images Figure 4.9 Detected optic disc and optic cup in fundus images

15 97 Figure 4.10 Neuroretinal rim area detection in fundus images 4.5 PERFORMANCE ANALYSIS FOR OPTIC CUP DETECTION i) To assess area of overlap between the computed region and ground truth of the optic cup, pixel-wise precision and recall values are computed using the Equations (4.8) and (4.9). Precision = TP TP + FP (4.8) Recall = TP TP + FN (4.9) where TP is the number of true positives, FP is the number of false positives and FN is the number of false negative pixels. Contours estimated by the ophthalmologist are set to be the ground truth. ii) Another method of evaluating the performance using F Score is given by the Equation (4.10). F = 2 Precision Recall Precision + Recall (4.10)

16 98 Value of F score lies between 0 and 1 and score will be high for an accurate method. Table 4.1 presents the quantitative assessment of the cup segmentation results using F score. Table 4.1 F score for cup segmentation Images Threshold Component analysis CMM CMM method achieves an average F score of 0.9 when compared to an average F score of 0.68 for thresholding and 0.74 for component analysis approach. The proposed cup segmentation method provides a significant improvement compared to the other two methods. 4.6 PARAMETER ESTIMATION FOR GLAUCOMA ASSESSMENT Determination of Cup to Disc Ratio The important feature is the cup to disc ratio (CDR), which specifies the change in the cup area. Due to glaucoma, the cup area will

17 99 increase slowly by intra ocular pressure and results in dramatic visual loss. Area ratio is selected to assess the overall segmentation accuracy achieved in all directions unlike the cup to disc diameter ratio which reflects the accuracy only in the vertical direction. Measuring the cupping area in the optic disc is an aiding tool for glaucomatous diagnosis as well as a following up tool for the glaucomatous optic disc to monitor the progression of the disease. CDR is one of the significant measurements that are found to be suspicious for glaucoma above a particular size usually 0.3.Increase in the cup-disc ratio or the enlargement of the cup over a period of time is diagnostic of glaucomatous disc damage. CDR is the ratio between the area of the optic cup and the area of the optic disc. CDR greater than 0.3:1 is the most often reported sign of disc damage.cdr > 0.3 indicates the suspection of glaucoma and CDR 0.3, is considered as a normal image as shown in Figure Area of the optic disc and cup is computed using the formula given below Area = ab (4.11) where a = major axis length of the disc boundary b = minor axis length of the disc boundary (a) CDR 0.3 (b) CDR > 0.3 Figure 4.11 CDR for normal and abnormal images Boundary of the detected cup compared with the threshold based approach and color component analysis are shown in Figure 4.12.

18 100 (a) Test images (b) Threshold based approach (c) Color Component analysis OPTIC DISC OPTIC CUP (d) Detection of optic disc using windowing technique and optic cup using color model. Figure 4.12 Disc and cup boundary extraction For left eye, boundary estimated by the CMM method and ground truth is closer at the nasal side. However, the boundaries at the nasal side are slightly different because there is no obvious cup structure on temporal side

19 101 and also the main blood vessels occlude on that region. For right eye, boundary is closer at the temporal side and slightly changes at the nasal side. Clinical and determined CDR values for few images are shown in Table 4.2. Table 4.2 Comparison of CDR values for few real time images Image CDR (Clinical) CDR (DW and CMM)

20 102 CDR is calculated from the fundus images using DW algorithm for the segmentation of optic disc and CMM technique for optic cup segmentation. To determine the performance of the approach, 40 retinal images are processed and their CDR are calculated. Images that are used as database are obtained from Aravind Eye hospital, Madurai. From the above analysis it is concluded that in few of the images there is a difference between the computed and ground truth values. Hence, there is a misclassification of two normal and two abnormal images and 90% accuracy is reported for 40 images. This is because the position of the optic cup in different sectors is not considered. CDR may be in the normal range but in the same image if cup position is closer to inferior or superior sector, there occurs severe rim loss and leads to loss of vision. CDR measure is inadequate to assess local cup changes and so sector-wise segmentation has to be performed Quantification of Neuroretinal Rim In order to quantify the focal cupping and to identify the stages of glaucoma, neuroretinal rim region is considered. Loss of axons in the eye leads to glaucoma which is reflected as abnormalities of the neuroretinal rim. Identification of the width of the neuroretinal rim in all sectors of the optic disc as shown in Figure 4.13 is of fundamental importance for the detection of diffuse and localized rim loss in glaucoma. Figure 4.13 Representation of structural features in fundus image

21 103 CDR is calculated regardless of the position of the cup and does not take into account the disc size. Optic disc is vertically oval and the optic cup is horizontally oval thus resulting in the characteristic shape of the neuroretinal rim. The normal optic disc usually have a configuration in which the Inferior (I) neuroretinal rim is the widest portion of the rim, followed by the Superior(S) rim and then the Nasal (N) rim with the Temporal (T) rim being the narrowest portion. Glaucoma frequently damages superior and inferior optic nerve fibers before temporal and nasal fibers, leading to thinning of the superior and inferior rims and violation of the rule. If CDR is less than 0.3 it appears to be a normal image, but if the cup region is closer to I or S, it is highly pathological and the patient suffers from visual damage due to severe neuroretinal rim loss. If CDR is greater than 0.3 and the cup region is closer to nasal or temporal quadrant, the patient is suspected to be in the early stage since rim loss has not occurred. So the calculation of neuroretinal rim area is essential for the accurate diagnosis of glaucoma NRRA detection algorithm Neuro Retinal Rim Area (NRRA) algorithm is described below 1. Boundary of the optic disc is extracted. 2. Optic cup segmentation is performed. 3. Binary image of the segmented disc and cup region is taken and cropped using a suitable mask. 4. Image is divided into four sectors (I, S, N, T) with adjacent lines making an angle of Calculate neuroretinal rim area in each of the sectors by calculating the number of white pixels in the rim region.

22 Neuroretinal rim area divided by the disc diameter is taken as ratio1.if ratio1 is greater than 0.2mm in the I and S region it refers to a normal image. 4.3 If ratio1 is less than 0.2mm in the inferior and in the superior region it refers to an abnormal image. For a large disc if (ratio1>0.3) disp('disk Damage - Stage 0a'); else if (ratio1>0.2 ratio1<=0.3) disp('disk Damage - Stage 0b'); else if(ratio1>0.1 ratio1<=0.2) disp('disk Damage - Stage 1 '); else if(ratio1>0.05 ratio1<=0.1) disp ('Disk Damage - Stage 2'); else disp ('Disk Damage - Stage 3'); end end end end

23 105 This process is repeated to calculate the rim to disc ratio in different images for left and right eye. Stage 0a and stage 0b refers to normal eye and ocular hypertension. Stage 1, stage 2 and stage 3 refers to early glaucoma, established glaucoma and advanced glaucoma. Neuroretinal rim area is the region between the optic disc and the optic cup. Rim area is measured in the inferior, superior, nasal and temporal quadrants. Usually the thickness of the rim area must be more in the superior and inferior regions when compared to the temporal and nasal regions in the normal images. To obtain the thickness in all the four quadrants shown in Figure 4.14, a binary image of the neuro retinal rim is taken as in Figure 4.15 and then cropped. A mask of the cropped image size is used to filter one quadrant. Then, the mask is rotated 90º to obtain the other quadrant area. Figure 4.15 shows the mask used for identifying rim area in the ISNT side of optic disc. Figure 4.14 Image with I S N T regions Figure 4.15 Binary image of the detected disc and cup

24 106 (a) Superior quadrant (b) Temporal quadrant (c) Inferior quadrant (d) Nasal quadrant Figure 4.16 Rim area detection using mask (a) Superior quadrant (b) Temporal quadrant (c) Inferior quadrant (d) Nasal quadrant Figure 4.17 Rim areas in ISNT sectors

25 107 This neuroretinal rim configuration gives rise to a cup shape that is either round or horizontally oval. Neuroretinal rim area is calculated by subtracting the area of the optic cup from area of optic disc. Normally the rim is the widest in the inferior temporal sector followed by the superior temporal sector, the nasal and the temporal horizontal sector. Neuroretinal rim area in each of the sector shown in Figure 4.17 is calculated by using a mask for each region and counting the number of white pixels in each of the I, S, N,T region. Large optic discs have large cups and often elevated CDR ratios. They can be incorrectly labeled as glaucomatous. Similarly small discs with a small CDR ratio may actually be pathologic and be erroneously classified as normal. ISNT rule states that rim width is greater in the Inferior region followed by superior, nasal and temporal regions. So rim width is calculated by calculating the number of white pixels in each of the sectors. Optic disc size varies from person to person and it could be a large, medium or small sized disc. If the neuroretinal rim width to the disc diameter is greater than 0.2 mm in the superior or in the inferior region it refers to a normal image. For pathological subjects, inferior and superior rim width thickness will be minimum which implies that neuroretinal rim width to a disc diameter is less than < 0.2mm. By calculating this ratio the extent of disc damage can be identified. Detection of rim width fails when there is an inaccurate segmentation of the optic disc and optic cup. Rim to Disc Ratio (RDR) in ISNT regions are mapped into a nomenclature which details the stage of the disease. 4.7 PERFORMANCE ANALYSIS OF STRUCTURAL FEATURES Discs are generally categorized as small, average and large. Small discs are less than 1.5mm, average size disc ranges between 1.5mm and 2mm and large discs are greater than 2mm. Stage 0a, stage 0b are the normal stages

26 108 and stages 1, 2, 3 denote the extent of disc damage for a pathological subject. For example when the rim to disc ratio ranges from 0.2 to 0.3, it refers to an abnormal stage 2 in a small disc, stage 1 in average size disc and stage 0b in a large disc. Table 4.3 Quantification of NRRA to identify stages of glaucoma Image I (µm) S (µm) N (µm) T (µm) Disc diameter (µm) Cup area (µm) CDR Rim to disc ratio (I) Rim to disc ratio (S) Stages of classifi cation a b a b a b a a a

27 109 Quantitative results for a few images are shown in Table 4.3. Asymmetry between Left Eye (LE) and Right Eye (RE) are shown in Table 4.4. Stages of disc damage in left and right eye are presented in Table 4.5.With the detected optic disc and cup boundary, disc diameter, cup area and cup to disc ratio are calculated. CDR does not consider the disc size. The stages, extending from no damage to far advanced damage are based on the width of the neuroretinal rim. In the first image shown in Table 4.3, CDR value greater than 0.3 indicates an abnormal image. In the same image, neuroretinal rim area is larger in the I region followed by S, N and T and hence ISNT rule is followed and resembles a healthy disc. Rim to disc ratios are then calculated for I and S region to identify the stage of the disease. Table 4.4 Rim to disc ratio in different images for left and right eye Images S. Left (µm) I. Left (µm) S. Right (µm) I. Right (µm) Disc diameter (Left) (µm) Disc diameter (Right) (µm) Rim to disc ratio (S.Left) Rim to disc ratio (S.Right) A A A Table 4.5 Stage of disc damage in left and right eye Images Rim to disc ratio(left) Rim to disc ratio(right) Stage of disc damage (Left) Stage of disc damage (Right) A A A The procedure is repeated for both left and right eye to assess the stage of disc damage. Based on neuroretinal rim area the extent of disc damage can be found in both left and right eye by taking into account the

28 110 asymmetry between left and right eyes. Stages of the disease can be found by calculating the rim loss in the inferior and superior quadrants. 4.8 CONCLUSION Optic cup segmented using CMM based on pallor provides a high rand index and lower displacement error when tested on a database of 40 images collected from Aravind Eye hospital, Madurai. Segmented results achieved an average F score of 0.9 compared to an average F score of 0.68 for thresholding and 0.74 in component analysis approach. Detected cup boundary using CMM technique showed good matching results with expert assessments of fundus images. As glaucoma progresses, the optic cup becomes larger and hence the cup to disc ratio becomes larger. Thus the CDR is higher for glaucoma subjects than for the normal subjects leading to the differences in the respective fundus images. Most of the previous studies on this discussion have addressed this issue by classifying the images based on cup to disc ratio. However CDR does not take into consideration the diameter of the optic disc nor does it directly describe the focal changes in the neuroretinal rim. There is a chance of misclassification in the images because large discs likely to have larger CDR ratio but with normal neuroretinal rims are more likely to be classified as glaucomatous, while small discs with small CDR ratio are more likely to be classified as normal whether they actually have glaucoma or not. Severe visual damage occurs when the cup region is very closer to the inferior or the superior quadrant. With the incorporation of disc size and rim width in clinical grading of the disc, rim to disc ratio reduces misclassification of images and identifies various stages of glaucoma by detecting rim loss in each quadrant. This feature also correlates strongly with the degree of glaucomatous visual

29 111 field damage. The estimated neuroretinal rim area to disc diameter based on the detected disc and cup boundaries showed good consistency when compared with the ground truth results. Experimental results showed that rim to disc ratio offers an excellent approach to distinguish between normal and glaucoma eyes and it outperformed CDR ratio. By categorizing discs as small, medium or large, the expectation of rim thickness is adjusted. Progression of the disease could be identified using the structural feature and can be used in Decision Support System (DSS) for the detection of glaucomatous optic nerves.