Explainable model for automatic melanoma detection using deep learning

Abstract

Melanoma is the deadliest form of skin cancer, which makes early detection essential for patients' survival. Whilst automated diagnostic systems have gained ground in dermatology, most are based on single-task architectures that have major clinical limitations: classifiers fail without flagging this when confronted with out-of-distribution lesions, whilst pure anomaly detectors flag anomalies without providing a diagnosis. In both cases, the models function as opaque 'black boxes', preventing clinicians from examining the reasoning behind their predictions. To address these gaps, this study proposes an explainable dual-branch system designed to aid clinical decision-making, which explicitly separates anomaly detection from lesion classification. The first branch is a Meta-CVAE, a conditional variational autoencoder adapted for single-class learning, trained exclusively on benign lesions and conditioned by patient metadata (age, sex and anatomical site), which quantifies the extent to which a lesion deviates from the normal appearance. The second branch is an EfficientNet-B4 classifier that classifies lesions into seven diagnostic categories. As each branch addresses only part of the clinical question, their results are harmonised by a joint clinical interpretation layer that identifies high-risk cases that a single model would have missed. To ensure the reliability of the explanations provided by the system, the explainability methods of each branch are systematically evaluated and ranked according to their causal fidelity. Evaluated on the HAM10000 dataset, the Meta-CVAE achieves a composite AUROC of 94.1% without any supervised exposure to malignant classes, whilst the classifier achieves an accuracy of 95.89%, a F1 score of 93.23% and a recall for melanoma of 92.73%. These results demonstrate that detection, diagnosis and interpretability can be effectively combined within a single framework that supports, rather than replaces, the dermatologist's clinical judgement.