An Explainable Deep-Learning Model to Aid in the Diagnosis of Age Related Macular Degeneration

Age-related macular degeneration (AMD) is the most frequent cause of blindness in people of advanced age. As AMD is asymptomatic in its early stages, this condition is normally identified in advanced stages of the disease, when treatments are less effective. To address this challenge, automated AMD image assessment systems offer the potential to significantly reduce the time, costs, and effort involved in screening. While previous works have demonstrated success for AMD detection using convolutional neural networks, their lack of explainability mechanisms limits their use in clinical settings. To address this limitation, we propose an explainable deep-learning approach using Local Interpretable Model-agnostic Explanations (LIME). Our model, based on RegNetY-320, achieved 86.5% accuracy, 85.21% sensitivity, and 91.01% specificity on the Automatic Detection challenge on Age-related Macular degeneration dataset. Through the LIME technique, we identified the specific areas in retinal images that influence the prediction of the model, providing a tool for clinical interpretation and enhancing diagnostic confidence.