AI-BASED DIABETIC RETINOPATHY DETECTION AND CLASSIFICATION USING RETINAL FUNDUS IMAGES AND TRANSFER LEARNING

Abstract

Diabetic retinopathy (DR) is a main cause of preventable blindness; however, early detection of DR is still difficult because of the subtle morphological changes in retinal fundus images. In response to this, we present a novel framework named SVG-DRNet (SVD-Guided Vision Transformer for DR Severity Grading and Lesion Region Exploration), which combines Singular Value Decomposition (SVD) based dynamic feature disentanglement with dual attention mechanisms for improved multi-scale feature extraction and fusion. SVG-DRNet first performs center-crop retina extraction, CLAHE enhancement and normalization operations on fundus images; then, it decomposes the main structural patterns from noise using the SVD method. A dual-attention learning module is then designed to combine features from both spatial and severity-grade layers, achieving both precise DR grading results for the APTOS 2019 dataset in 5 classes and interpretability in terms of lesion regions. The extensive experiments show that SVG-DRNet outperforms the custom CNN baseline, VGG16 baseline and ResNet50 baseline in terms of validation accuracy (92.1%) and macro F1-score (91.1%). The system not only promotes the development of clinical level DR screening but also highlights the clinically relevant areas of the lesion, which can facilitate timely treatment in limited-resource countries.

Keywords :

Diabetic retinopathy, SVD-guided framework, Dual attention learning, Fundus imaging, Transfer learning, APTOS 2019, Lesion region exploration