An Explainable Deep Learning Framework for Medical Imaging Diagnostics: Integrating Grad-CAM, SHAP, and Attention Mechanisms

Medical imaging diagnostics has reached impressive accuracy through the incorporation of deep learning model; however, deep learning as black-box classifiers hinders its clinical use. In this paper, we propose a new explainable deep learning model to combine Gradient-weighted Class Activation Mapping (Grad-CAM), Shapley Additive explanations (SHAP) and attention models into a single explainable framework to enable both single and multiple layer explanation for medical image analysis. Our framework features visual heatmaps, quantitative feature attributions, and inherent attention-based interpretability across the "diagnostic" – "clinical explainability" divide. The proposed framework localization accuracy of 94.7% and faithfulness score of 0.92 further highlights the high level of explainability of the results achieved and evaluated on ISIC 2020 skin lesion dataset (25,331 im-ages) and Kaggle Diabetic Retinopathy (DR) dataset (28,101 images). By incorporating uncertainty quantification via Monte Carlo Dropout, the decision of whether to trust the model's prediction is done in a better way by identifying the uncertain ones. This study shows that our hybrid explainability method does not fall short of single method explainability techniques and offers the full coverage of making explainability comprehensive enough to align with radiological assessment patterns. Overall, this work advances the development of trustworthy and regulatory-compliant AI systems in medical diagnostics, enhancing the alignment of algorithmic effectiveness with clinical applicability.