Ante-hoc Framework for Interpretable Skin Lesion Classification

Despite high reported accuracy in automated dermatological screening, many deep learning ensembles remain limited by opaque architectures and reliance on post-hoc explainability methods. This paper proposes a lightweight ante-hoc semantic framework optimized for low-latency edge execution to minimize bandwidth overhead in decentralized clinical networks. The architecture uses a knowledge-distillation pipeline to transfer representations from a foundation model to a computationallyefficient student model. An ante-hoc concept bottleneck model (CBM) architecture constrains the network to predict eight dermoscopy-derived visual concepts before producing the final lesion label. On HAM10000, an anonymized public dataset, the framework achieves an overall accuracy of 90.75%, a balanced accuracy of 84.55%, and an area under the curve (AUC) of more than 0.9500, suggesting that interpretability can be introduced with limited loss in discriminative performance. The resultingconcept-based predictions are auditable and could support clinicians in prioritizing high-risk cases, subject to further clinical and deployment studies.