An Explainable Deep Learning Framework for Skin Cancer Diagnosis with Uncertainty Quantification on a Multi-Institutional Dataset

Abstract

Deep-learning-based diagnosis of skin cancer is now at the dermatologist level, yet further challenges remain in its application in practice: lack of generalization between institutions, lack of confidence in estimates, and lack of transparency in decision-making. This paper presents and tests an explainable deep learning model that utilizes a Bayesian uncertainty quantification to facilitate automatic skin cancer diagnosis from multi-institutional dermoscopic images. The framework combines data harmonization, transfer learning for convolutional features extraction, Monte Carlo Dropout for uncertainty estimation, and Grad-CAM for visual explainability, enhancing the robustness, reliability, and interpretability. A comprehensive experimental evaluation showed that the proposed approach outperforms CNN baselines in terms of accuracy, sensitivity, specificity, and AUC, and that the performance drop with cross-institutional domain shift was less than that of CNN baselines. Uncertainty-sensitive deferral strategies enhanced confidence in the diagnosis of high-confidence cases, and calibration analysis showed good calibration in the probabilistic prediction. Visual attribution analysis confirmed stable patterns of attention to lesions, ensuring an agreement between clinicians' beliefs and transparent human-AI interaction. To conclude, the results indicate the need for a common framework for explainability, uncertainty quantification, calibration, and cross-institutional testing as a crucial step toward clinicians' safety and reliability when using AI for the diagnosis of skin cancers.