Attention-Guided Swin Transformer for Retinal Disease Classification in Fundus Images Using Self-Distillation Mechanism

Abstract

Organizing systems which detect retinal diseases in fundus images face difficulties because of the existence of minute visual details alongside overlapping medical patterns. This research develops an Attention-Guided Swin Transformer (AG-SwinT) framework to improve feature extraction through spatial attention functions and maintain local information retention. We developed an Attention-Guided Swin Transformer (AG-SwinT) system which unites a modified Swin Transformer with an attention-guided module for fundus image region-specific weighting selection. Our design employs self-distillation because information flows from teacher models to student models with identical structures thereby strengthening the learning of important visual features. The developed network received training and validation using a large dataset which contained 9,400 fundus images labeled with five typical retinal pathologies. AG-SwinT reached a classification accuracy rate of 94.7% for all disease groups which exceeded regular CNN methods (89.3%) and standard Swin Transformer approaches (92.1%). Evidence showed that both sensitivity and specificity measured at 93.8% and 95.2% respectively. The most remarkable enhancements in performance occurred in early-stage diabetic retinopathy detection with subtle symptoms because the new model surpassed baseline accuracy by 15.4%.