Explainable Transformer-Based Retinal Disease Classification.
Abstract
Retinal diseases are a leading cause of visual impairment globally. Early detection is crucial for effective management, and recent advances in ocular imaging and deep learning have facilitated the development of automated diagnostic systems. Traditional deep learning methods often struggle to effectively capture the overall structural context of the retina and the subtle, localised pathological features required for accurate multi label classification. Moreover, their decision-making processes typically lack transparency, limiting clinical trust and adoption. In this study, we propose an explainable transformer-based fine-tuned model for retinal disease classification that uses a Vision Transformer architecture enhanced with a Multiple Instance Learning (MIL) mechanism. The model is fine-tuned on the multi label retinal disease (MuReD) dataset, enabling it to simultaneously predict multiple co-existing pathologies. Experimental evaluation on the MuReD dataset demonstrates state-of-the-art performance, with the proposed fine-tuned MIL-VT achieving an ML_F1 (multi label F1) score of 0.672, representing an approximately 10 percentage-point increase over recent transformer-based baselines, alongside higher ML_mAP (multi label mAP) and comparable AUC scores. Additionally, we evaluate our model's performance on an independent Eye Disease Image (EDI) dataset, where it maintains better accuracy, achieving a macro-average F1 of approximately 0.92, highlighting its robust generalisation across datasets. An ablation study further validates the effectiveness of the MIL head and selective fine-tuning strategy, showing their substantial contributions to performance gains. To enhance clinical interpretability, we incorporate Grad CAM visualisations that highlight pathological regions associated with each predicted condition. The combination of better classification performance, explainability, and cross-dataset robustness underscores the potential of our proposed framework to improve automated retinal disease screening in diverse clinical settings.