A leakage-aware evaluation framework for reliable deep learning-based melanoma detection

Abstract

Deep learning has shown promising performance for automated melanoma detection; however, many published studies still rely on image-level data splits that can introduce information leakage and produce overly optimistic estimates of real-world clinical performance. This study aims to develop a leakage-aware, reliability-centered evaluation framework for deep learning–based melanoma detection from dermoscopic images that better reflects realistic deployment conditions. The study applies lesion-level GroupKFold cross-validation to a 15,000-image natural-prevalence cohort and compares two transfer-learning baselines, MobileNetV2 and EfficientNetB0, under matched preprocessing, training, and evaluation settings. Beyond conventional discrimination performance, the framework assesses probability reliability using Expected Calibration Error and Brier Score, examines the effects of leakage-free post-hoc temperature scaling, and evaluates uncertainty-aware selective prediction through Monte Carlo Dropout. The findings show consistent discrimination performance, with a mean AUC around 0.82, together with strong reliability characteristics, including an Expected Calibration Error of approximately 0.028 in the primary natural-prevalence evaluation. The results further indicate that uncertainty-based selective prediction can substantially improve safety by restricting automated decisions to higher-confidence cases, reducing the error rate to about 6.7% when retaining the most confident 70% of predictions. Overall, these findings have practical implications for trustworthy clinical AI by showing that leakage-aware validation, empirical calibration analysis, and uncertainty-guided deferral mechanisms are important for safer, more clinically meaningful, and more deployable melanoma screening systems.