Abstract
Diabetic retinopathy (DR) is a leading cause of vision impairment, requiring accurate early detection to prevent irreversible blindness. This study proposes DiabFoRMaxNet, a multimodal deep learning framework for optimized DR detection and classification using retinal fundus and optical coherence tomography angiography (OCTA) images. The framework integrates five major stages: preprocessing, feature extraction, feature segmentation, feature optimization and classification. Preprocessing combines Z-score normalization, green channel extraction, adaptive histogram equalization and bilateral filtering to enhance retinal image quality. AlexNet is employed for hierarchical feature extraction, while Fully Convolutional Networks (FCN) perform lesion-aware pixel-wise segmentation. The Fossa Optimization Algorithm (FOA) refines extracted features by reducing redundancy and RankMax classification improves robustness under class-imbalanced conditions. Using an augmented multimodal dataset expanded from 222 original images under controlled augmentation protocols, DiabFoRMaxNet classifies NO DR, MILD DR and MODERATE DR categories with high precision. Experimental evaluation demonstrated superior performance over conventional deep learning models, achieving 99.18% accuracy for retinal fundus images and 99.69% accuracy for OCTA images, along with strong sensitivity, specificity, precision and F1-score. The proposed multimodal architecture enhances lesion localization, computational efficiency and diagnostic reliability, indicating strong potential for scalable AI-assisted early DR screening and clinical ophthalmic decision support.
Keywords: Diabetic Retinopathy, DiabFoRMaxNet, OCTA, RankMax ClassiGier, Retinal Fundus.