Explainable retinal disease classification and localization through Convolutional Neural Networks

Retinal diseases pose significant challenges to vision globally, affecting a substantial portion of the population. The reliance on expert clinicians for interpreting Optical Coherence Tomography images underscores the need for automated diagnostic process. In this paper, we propose a method aimed at automatically detecting and localizing retinal disease through deep learning convolutional neural networks starting from the analysis of optical coherence tomography imaging. In detail, we propose and design a novel deep learning model, i.e., FCNNplus, for the classification task of retinal disease, reaching 93.3% in accuracy. The focus is not only on achieving a satisfying retinal disease diagnosis but also on emphasizing the role of CAM algorithms in localizing disease-specific patterns to propose a method considering the explainability and reliability behind the prediction. FCNNplus reports precise and accurate heatmaps localization, correctly identifying the presence of the retinal disease in the images. We take into account an index of similarity aimed to enhance the qualitative aspects and provide a measure of the visual explanation coming from the heatmaps (i.e. the areas of the image under analysis that, from the model point of view are symptomatic of a certain prediction). • Paper focus: explainable AI for medical diagnosis using CAM heatmaps for trust. • FCNNplus: deep learning for retinal disease classification from OCT, 93.3% accuracy. • FCNNplus reports optimal heatmaps for the qualitative point of view. • IM-SSIM index enhances heatmap quality, improves AI visual explanation reliability.