What’s Missing in Medical XAI? EIGradCAM for Refined, Faster, and More Reliable Explanations

Deep learning (DL) has significantly improved medical imaging, enabling automated diagnosis and segmentation. However, its black-box nature limits clinical adoption, necessitating Explainable Artificial Intelligence (XAI) solutions for reliable decision-making. Class Activation Mapping (CAM)-based methods namely Grad-CAM and Grad-CAM++ offer explainability but suffer from coarse localisation, modality-specific inconsistencies, and high computational costs. To overcome these limitations, we propose Efficient Integrated Grad-CAM (EIGradCAM), an XAI-CAM method that integrates gradient accumulation and morphological refinement resulting in an improved spatial precision and computational efficiency. Evaluations on MRI and ultrasound datasets demonstrate EIGradCAM’s superiority in localising pathological regions with higher Dice Coefficient (DC) and Jaccard Index (JI) scores, indicating enhanced performance accuracy, and lower Hausdorff Distance (HD) values, signifying improved boundary adherence. Additionally, EIGradCAM reduces inference time by 34.08% over Grad-CAM and 51.65% over Grad-CAM++, making it suitable for real-time AI-assisted diagnostics. Its efficiency is achieved by refining attribution maps without relying on computationally expensive higher-order derivatives. Our findings establish EIGradCAM as a robust and computationally efficient XAI technique, bridging the gap between DL explainability and clinical usability.