Title: Enhancing Robustness in Skin Lesion Classification Through Contrastive Meta-Domain Adaptation Across Clinical and Acquisition Contexts

Abstract: The deployment of deep learning models for dermatological image analysis is often hindered by their vulnerability to variations in image acquisition and domain-specific visual traits, which frequently result in diminished performance within real-world clinical environments. This study explores the impact of domain shifts and visual artifacts on the efficacy of deep learning approaches for skin lesion classification. To address these challenges, we introduce an adaptation framework centered on the concept of visual meta-domains. This strategy facilitates the transfer of visual representations from extensive dermoscopic datasets to clinical image domains, thereby bolstering generalization capabilities. Our experimental evaluations, conducted across several dermatology datasets, demonstrate uniform improvements in classification accuracy and a notable reduction in the performance disparity between dermoscopic and clinical imagery. These findings underscore the critical role of domain-aware training methodologies in developing robust, deployable systems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC