Multi-Classification of Non-Proliferative Diabetic Retinopathy Through Integrated Machine Learning Approach in Fundus Images

Abstract

Diabetic Retinopathy is an ocular sickness resulting in the visual disability, the treatment and cure of this eye disease becomes comfort if the disease is identified at the earliest. The present study conceives an integrated machine learning approach for the multi-level multi-classification of the earliest stage of diabetic retinopathy called Non-Proliferative Diabetic Retinopathy. At the first level, the disease features are classified and at the second level, the disease severities are classified. The implementation of the work ensues with the fundus images undergoing grayscale conversion and median filter for preprocessing. Then, the statistical feature vectors like local binary patterns, histogram of gradients, and gray level co-occurrence matrix are extracted and fed into a multi-class support vector machine for classifying the non-Proliferative diabetic retinopathy disease features called microaneurysm, intra-retinal hemorrhages, and hard exudates. The classified features are classified into non-proliferative-diabetic-retinopathy disease severities namely mild, moderate and severe with the k-Nearest neighbor, random forest, and naive bayes methods. The proposed classifiers are assessed and validated in terms of accuracy and execution time; comparatively the k-Nearest neighbor classifier achieved a better result of 99% accuracy and the least processing time.