Robust and Equitable Skin Lesion Classification via Adaptive Feature Alignment and Structured Sparse Interleaved Learning

Abstract

Automated dermoscopic image classification remains challenging because of substantial interdataset distribution shifts, severe class imbalance, and high intra-class variability. This study introduces a computationally efficient and diagnostically reliable framework that integrates two novel components: an Adaptive Feature Alignment Module (AFAM) for prototype-guided refinement of pretrained feature embeddings, and a Sparse Interleaved Feature Module (SIFM) that applies structured sparsity within fully connected transformations to yield compact and discriminative representations. Comprehensive evaluations on six public benchmarks, including comprehensive external validation, demonstrated strong generalizability and consistently superior performance relative to diverse baseline and state-of-the-art models while incurring only negligible computational overhead. Quantitative Grad-CAM analysis confirms high spatial correspondence between activation maps and clinically relevant lesion regions, and fairness assessments reveal minimal subgroup-level performance disparities. Ablation studies further substantiate the complementary contributions of AFAM, SIFM, and imbalance-aware optimization. Overall, the proposed alignment–sparsity paradigm provides a robust, generalizable, and interpretable framework for dermoscopic image classification with strong potential for equitable point-of-care clinical deployment. © 2017 IEEE.