Multi-Label Retinal Screening System for Disease Detection with Hybrid CNN

Abstract

Retinal disease diagnosis using fundus imaging is a critical task in ophthalmology. Traditional deep learning approaches often struggle with uneven illumination, small lesion detection, and multi-label disease co-occurrence. This paper presents an upgraded hybrid framework integrating disease-aware illumination normalization, multi-scale lesion-sensitive CNN, Vision Transformer (ViT), bidirectional LSTM (BiLSTM), attention refinement, and label-dependency guided prediction. Experimental evaluation on a large-scale dataset of 249,620 fundus images annotated for 39 retinal diseases demonstrates superior performance compared to previous CNN–ViT–LSTM models, achieving a macro-F1 score of 0.944 and reducing hamming loss to 0.018. These results confirm the clinical robustness of the proposed methodology.

The upgraded framework extends the earlier hybrid CNN–ViT–LSTM retinal disease diagnosis system by incorporating three major methodological improvements: disease-aware illumination normalization, multi-scale lesion-sensitive feature extraction, and label-dependency guided multi-label prediction. The overall objective of the upgraded design is to improve robustness against fundus image quality variations, enhance sensitivity toward small retinal lesions, and improve the recognition of co-existing ophthalmic conditions.