A NOVEL APPROACH TO GLAUCOMA SCREENING USING OPTIC NERVE HEAD THROUGH IMAGE FUSION AND FRACTAL GEOMETRY

The Glaucoma is a typical eye issue that causes vision loss. It leads to visual impairment if it is untreated on time. Normally, vision loss is slow and not perceptible. Regular and systematic eye assessments are suggested for persons from middle age to prevent further vision loss. The proposed system introduced a new technique in the field of ophthalmology to diagnose glaucoma in an effective way using Image Fusion and Fractal Geometry techniques. The optic cup and disc are extracted from fundus images using K-means and Thresholding techniques. Optic cups generated by the techniques are combined to obtain the better cup region using the image Fusion Technique to improve the glaucoma screening process. The same process is applied on optic discs to obtain the fused disc area. The Box counting fractal dimension estimation technique from Fractal Geometry is applied on fused areas to classify the image as either healthy or glaucoma. Results of these two techniques are evaluated on a publically 4285 A NOVEL APPROACH TO GLAUCOMA SCREENING available HRF dataset and obtained the accuracy of 97%.


INTRODUCTION
Billions of people continue to open increased risk of visual deficiency or huge visual disability. It reduces the quality of human life if the condition is undiscovered, left untreated and not diagnosed on time. Glaucoma is the one cause for visual deficiency. Hence, early discovery of glaucoma is important to strengthen the assistances of treatment and it remains a challenge when it is undiagnosed in the community. According to survey, it is shown that expenses of treating glaucoma in many of developed countries increase as severity of glaucoma [1]. If glaucoma patients are not correctly identified and treated at the prior stages leads to reduction in reserve funds and causes burdens on healthcare facilities. In current era, medical field equipped with advanced instruments but require new approaches to diagnose diseases in its initial-phase. There are different clinical parameters and approaches to detect glaucoma in the early stage [2][3][4][5][6].
Following sections describes the different methods proposed by various authors to detect glaucoma.
Ajesh et al., [7] developed a methodology for finding glaucoma condition at the initial stage by examination of retinal features extracted from fundus image using imaging process techniques.
Authors presented an improved machine learning algorithm to discover the disease. A discrete wavelet transform (DWT) is used for classifying the diseases. Proposed methodology provided the better results and achieved 95 percentage of accuracy. J. Carrillo et al., [8] illustrated the basic concepts of glaucoma and different screening techniques. This work presented the computational tool to extract disc and cup areas from the fundus images using thresholding technique. The proposed method was tested on Center of Prevention and Attention of Glaucoma in Bucaramanga, Colombia dataset and obtained the accuracy of 88.5%.
Guangzhou An et al., [9] developed an convolutional neural network (CNN) and random forest (RF) algorithms based technique to diagnose the open-angle glaucoma. Work is focused on retinal nerve fiber layer thickness; optic disc and macular ganglion cell complex (GCC) present in Optical coherence tomography and fundus data. Proposed method generated deviation and thickness maps using segmentation approaches. It is evaluated on 357 images and obtained an AUC of 0.963. In [10], authors presented a new tool for detection of glaucoma using fundus images. They extracted local configuration pattern (LCP) based features and texton from the images to analyze the glaucoma status. Texton are generated by applying adaptive histogram equalization and convolution operations on images. The proposed method achieved the accuracy of 95.8%. Simonthomas, S et al., [11] presented new computerized approach to glaucoma diagnosis. The authors extracted the Gray Level Co-occurrence Matrix (GLCM) and Haralick based texture features to diagnose the glaucoma disease. The Image pre-processing is performed to eliminate the noises. Later, Noise free image is used to extract the GLCM and thirteen Haralick texture features. Finally, all these features are feed to k-nearest neighbors classification technique and achieved the 97% accuracy.
Abhishek Pal et al., [12] presented auto encoding system called G-EYENET to identify the glaucoma. The modified u-net CNN is used to extract Region of Interest consisting of Optic Disc from fundus images. Authors considered RIM ONE v3, Drishti-GS, HRF databases for training the system and DRIONS-DB for testing purpose. The G-EYENET achieved the AUC of 0.923.
Juan J et al., [13] developed a transfer learning approach to glaucoma detection. Optic disc is segmented by morphological operations then VGG-19 net is used for transfer learning. The approach achieved the AUC of 0.94. Alan Carlos de Moura Lima et al., [14] proposed a CNN based RNN architecture version 50 to study process of glaucoma in patients. The approach achieved the accuracy of 90%. Annan Li et al., [15] presented an approach to glaucoma detection using CNN. The deformable shape model is used to segment the optic disc from image. From

METHODOLOGY
The steps involved in the proposed method are illustrated in Figure 1. Color image captured from the fundus camera is input to proposed method. It works on two methods 1) simple thresholding and 2) K-means clustering techniques. In simple thresholding, The RGB channels are separated from the color image. Green channel image is selected for further processing because it has high density vascular architecture at optic nerve head region. The region of interest (RoI) considered for glaucoma detection is optic nerve head (ONH) area which is extracted by simple thresholding method. Reason for this RoI selection is ONH has more damaged area in glaucoma condition compared to healthy.

B. K-means clustering
K-means clustering is a segmentation algorithm. It divides the image into different clusters, each cluster consisting pixels similar to other pixels in the same cluster and different than those in other clusters. Algorithm divides the image into K clusters, among these clusters one of the clusters represents the disc and one more represents the cup. By experimental observation, value of K selected is 4. The cluster 3 represents the disc area and cluster 4 represents the cup area.
Working principle of K-means algorithm is as follow

Algorithm: K-means
Input: color image (I), K=4 Output: Segmented binary Images(c1,c2,c3,c4) Step1: Randomly select K pixels as initial clusters Step 2: Allocate each pixel in the image to the closest centroid Step3: Calculate the center of the clusters Step 4: For every clusters, Find the distance between pixels and centers using Euclidean distance.
Step 5: Based on calculated distance reassign the pixels to nearest clusters Step 6: Again find the center of new clusters Step 7: Repeat the steps 4,5 and 6 until pixels don't change the cluster

C. FD Estimation
Fractal dimensions are used to define the dimension of asymmetrical, irregular objects. The best widespread technique to estimate the irregularities in objects is Counting Method [17].

RESULTS
The proposed approach is evaluated on publically available High-Resolution Fundus (HRF) Database found at webpage [18]. It has 45 retinal fundus images, out of which 15 healthy images and 30 are glaucoma affected. In this work, one ophthalmologist marked optic disc and cup areas are considered for accuracy calculation. Dice method [19] is not a sufficient method to measure the performance of proposed method because area marked by the ophthalmologist not accurately correlating with area extracted by the proposed method due to high level of pixels variations in glaucoma images. Hence, both dice coefficient and accuracy estimation [20]  Steps involved in the glaucoma image processing are illustrated in the figure 2. Figure 2 a) illustrates the structure of retinal fundus image [21]. b) Shows healthy retinal image. c) Represents the glaucoma image. d) Illustrate the RoI extracted from the fundus image using simple thresholding method with a threshold value of 170. e) Illustrates the optic cup extracted from optic nerve head using simple threholding method with a threshold value of 150 after obtaining the RoI of size 252X252. f) Represents boundaries extracted using morphological operations. g) illustrates the FD calculation using Box counting method, the x-axis represents the log(r) value (the number of boxes in the vertical grid) and y-axis represents the log N(r) value (number of boxes covered the cup boundaries) and obtained FD value using (2) is 1.0438.    [7] DWT 95% Carrillo et al., [8] Thresholding 88.5%.

97%
Abhishek Pal et al., [12] G-EYENET AUC of 0.923 Juan J. et al., [13]  glaucoma images as glaucoma. Therefore accuracy achieved is 97%. Table 2 illustrates the comparison of accuracy of method with existing approaches and obtained the better accuracy compared to other approaches. Figure 6 illustrates the Calculated FD values using thresholding, K-means and image fusion techniques on HRF dataset. Threshold FD value for healthy cup is below 1.035 and above 1.035 is considered as glaucoma. Threshold FD value for healthy disc is below 1.32 and above 1.32 is considered as glaucoma. Figure 6 illustrates

CONCLUSION
Early detection and diagnosis of glaucoma in everyday practice is more essential in order to ensure potential benefits for early treatment. In this work, new technique is developed to improve glaucoma evaluation. Calculating the accurate cup and disc areas using single technique in glaucoma evaluation is a challenging task. Thus, outputs of two feasible methods (thresholding and k-means) are combined to obtain accurate cup and disc areas, which in turn supports better glaucoma identification. A novel parameter called Fractal Dimension is calculated on cup and disc areas using box counting technique. Results show the proposed system is efficiently classified the images. This illustrates the image fusion technique is better approach to improve the accuracy of glaucoma detection. The proposed method is tested on HRF Dataset and provided better outcomes as compared to existing approaches. The results obtained are correlated with the results of the ophthalmologist and provided accuracy of 97%. Therefore, the proposed system can be used for early detection of glaucoma and as the decision support system for ophthalmologists.

CONFLICT OF INTERESTS
The author(s) declare that there is no conflict of interests.