Diabetic Retinopathy Detection Using MSLBP, Harris Corner Detection, and SVM Classifier

Abstract

Diabetic retinopathy, a condition affecting diabetic patients, involves the formation of clots, lesions, or hemorrhages in the retina's light-sensitive area. Elevated blood sugar levels lead to vessel blockages and the growth of new vessels, forming mesh-like structures. Accurate evaluation of the retinal vasculature is essential for effective diagnosis. In assessing diabetic retinopathy, fundus scans undergo preprocessing and segmentation, including image enhancement, retinal mask extraction, blood vessel segmentation, optic disk extraction, and lesion candidate region extraction. Branching blood vessels are extracted using thresholding, followed by adaptive histogram equalization and morphological opening to enhance images and remove falsely segmented areas. The proliferation of optical nerves is notably higher in diabetic patients. Additional features are extracted from lesion candidates using a hybrid approach of Harris corner detection and Multi-Scale Local Binary Pattern (MSLBP). A random forest classifier is employed to classify diabetic retinopathy presence. Performance evaluation is conducted on two datasets: DIARETDB1, a standard dataset, and a medical institution's dataset comprising fundus scans of both affected and normal retinas. Experimental results demonstrate the proposed algorithm's superiority over traditional methods, achieving 98.7% accuracy and 97.2% precision on the DIARETDB1 dataset.