Enhancing Interpretability in Retinal OCT Analysis Using Grad-CAM: A Study on the AROI Dataset

The increasing use of deep learning in medical imaging, particularly optical coherence tomography (OCT), has transformed retinal disease diagnosis and segmentation. However, the limited interpretability of these models remains a significant barrier to clinical adoption. This study employs Gradient-weighted Class Activation Mapping (Grad-CAM) to enhance the interpretability of U-Net-based architectures for OCT image analysis. Using the Annotated Retinal OCT Image (AROI) dataset, we evaluate U-Net and Attentionbased U-Net architectures, comparing their performance in segmenting retinal layers and pathological fluids. Grad-CAM generates visual explanations that highlight regions in OCT images influencing model predictions. Qualitative analysis shows that heatmaps from the Attention-based U-Net align more closely with clinically relevant features, especially in cases with severe pathological changes. Quantitative evaluation demonstrates improved segmentation performance, with Weighted Overlap scores confirming the positive impact of attention mechanisms on diagnostically critical regions. By integrating interpretability into segmentation workflows, this study bridges the gap between AI models and their clinical application, fostering transparency and trust in retinal disease diagnostics.