An efficient automated system for detection of diabetic retinopathy from fundus images using support vector machine and bayesian classifiers

Abstract

The preliminary signs of diabetic retinopathy include micro aneurysms, haemorrhages and exudates. Early diagnosis and timely treatment can prevent vision loss in patients with long term diabetes. In this paper we used two algorithm based on filtering operations, morphological transformation and region growing method to extract features for detection of micro aneurysms, haemorrhage and non linear diffusion segmentation followed by colour histogram based clustering techniques is used to differentiate hard and soft exudates. Experimental evaluation of the algorithm has been done with images collected from Deepam Eye Hospital, Chennai, Tamilnadu, India and a database consisting of 77 abnormal and 20 normal images was created. In addition performance of the proposed algorithm is also verified on the publically available DIARETDBO database. Based on the features obtained, each image is classified as normal or abnormal with Support Vector Machine, BayesianNetwork. Classification rate of 95% is obtained with SVM and 90% with Bayesian classifier.

Review of Preprocessing Techniques for Fundus Image Analysis

“…Without the use of any contrast enhancement procedure optimal AM-FM results in true candidates extraction to prove as reliable algorithm to not only enhance the low intensity of the images but also extraction of multiple features to detect actual candidates through classification in [127][128]. First white top hat transforms the geodesic detection followed by color histogram thresholding is the sequential steps provided in preprocessing module by [129]. Goatman et al [130] proposed a method for contrast enhancement in which the image contrast stretching was used to cover full pixel dynamic range excluding the dark surrounding border pixels for normalization.…”

Image processing and classification in diabetic retinopathy: A review

“…Vlachos et al [9] utilized the model used by Forachhia et al [8] for background normalization. In [10], Narasimhan et al filtered the green plane image by median filter and divided green plane image by median filtered image for normalization. Goatman et al [11] proposed a method for contrast enhancement in which the image contrast stretching was used to cover full pixel dynamic range excluding the dark surrounding border pixels for normalization.…”

“…As a preprocessing step, color normalization and contrast enhancement were done. Narasimhan et al [10] proposed adaptive thresholding and region growing in combination with Bayesian and SVM classifiers for the MAs classification. In the preprocessing step green plane extraction, normalization of image by median filter, white top hat transform and geodesic dilation were utilized.…”

“…Whereas in [30] Lochan et al proposed canny operator in combination with Kirsch's algorithm and classical matched filter for the feature extraction of exudates in retinal image. Narasimhan et al [10] employed white top hat transform, geodesic dilation, color histogram thresholding and region growing in combination with Bayesian and SVM classifier.…”

Hierarchical Detection of Hard Exudates in Color Retinal Images

“…This results from the existing of interference such as other light lesion regions that have same intensities as HEs. After defining some inherent features such as geometry, color, contrast and texture features, many machine learning methods such as k-nearest neighbor classifier and linear discriminant classifier [11], statistical classification [12], neural network [13], [14], Bayesian classifiers [15] and SVM [16] can used to classify the true HEs.…”