Interpretable Deep Learning for Retinal Disease Detection

Abstract

Eye diseases are a leading cause of global vision impairment, yet detection is often hindered by limited resources and geographic barriers. While Deep Learning (DL) offers automated diagnostic solutions, its black-box nature challenges clinical trust. This research investigates whether interpretability can be integrated into DL models for eye disease classification without compromising performance. Using an EfficientNet-B0 architecture, two models were trained on fundus images: one trained only on real images and another incorporating synthetic images generated by a Deep Convolutional Generative Adversarial Network (DCGAN) to simulate insufficient data. Both training approaches were evaluated using 3-fold cross-validation and analysed with eXplainable Artificial Intelligence (XAI) techniques: Grad-CAM for visual heatmaps and SHAP for feature-contribution insights. A General Practitioner (GP) also assessed the clinical relevance of selected Grad-CAM outputs. Cross-validation produced mean accuracies of 91.59% ± 0.81% for the real-only model and 90.51% ± 0.83% for the mixed model, with no statistically significant difference in accuracy (paired t-test, p = 0.366). Synthetic image quality was further evaluated using fidelity and diversity metrics, while additional analyses examined classification errors, calibration, and robustness to explanations. These results demonstrate the feasibility of integrating interpretability with DL models for retinal disease classification, including scenarios involving synthetic data augmentation, supporting further development of transparent and accurate automated medical diagnostic tools.