K-Means Clustering Algorithm for exudates segmentation in fundus images for the diagnosis of diabetic retinopathy

Abstract

Diabetic retinopathy currently remains one of the major causes of blindness or loss of vision in people. Automatic detection of diabetic retinopathy has been one of the challenging research fields in recent history. Exudates, optic disc, hemorrhages, cotton wool spots, and microaneurysms can slow the development of an illness if diagnosed early and with adequate treatment. The monitoring and identification process of the lesions associated with diabetic retinopathy remain very tedious tasks loaded with repetition and are prone to errors. It is in this regard that this research proposes a new methodology for the diagnosis of diabetic retinopathy in an automated way. First of all, EYEPACS-1 and APTOS 2019 retinal images datasets are obtained. After fundus image quality collection, some enhancements are made through adaptive histogram equalization and the Haar discrete wavelet transform with the support of filtering based on a Gaussian matched kernel. Further, the fovea and blood-retinal vessels are eliminated; through the K-means clustering segmentation algorithm, the areas of an optic disc are highlighted by the Gaussian blur approach. Color histogram features are then extracted from these segmented areas before being fed into a deep learning model, CNN, to classify healthy subjects versus those affected by DR with severity grading. According to the simulation findings, the accuracy posted by the proposed model was 97.25% against the EYEPACS-1 dataset and 96.85% against the APTOS 2019 datasets & samples.