Hierarchical severity grade classification of non-proliferative diabetic retinopathy

Abstract

Curability of diabetic retinopathy (DR) abnormalities highly rely on regular monitoring, early-stage diagnosis and timely treatment. Detection and analysis of variation in eye images can help the patient to take the early action before progression of the disease. Vision loss can be effectively prevented by automated diagnostic system that assist the ophthalmologists who otherwise practice manual lesion detection processes which are tedious and time-consuming. This paper proposes a hierarchical severity level grading (HSG) system for the detection and classification of DR ailments. The retinal fundus images in the proposed HSG system are categorized as grade 0 (indicating Non-DR class) and DR severity grades 1, 2, 3 depending upon the number of anomalies; microaneurysms and haemorrhages in the fundus images. The challenge of retinal landmark segmentation, DR retinal discrimination and DR severity grading have been addressed in this work contributing to the novelty of the proposed approach. For non-DR and DR classification, the proposed system achieves an overall accuracy of 98.10% by SVM classifier and 100% by kNN classifier. Hierarchal discrimination into further grades of abnormalities resulted in accuracy values of 95.68% and 92.60% with SVM classifier using Gaussian kernel and, 97.90% and 95.30% employing fine kNN classifier. The HSG system demonstrates a clear improvement in accuracy with significantly less computational time comparative to the other state-of-the-art methods when applied to the MESSIDOR dataset. IDRiD dataset is also evaluated for performance validation of the proposed HSG system yielding a maximum of 94.00% classification accuracy using a kNN classifier with a computational time of 0.67 s.