Answers are grounded in this paper and cite their sources.
Automated multiclass skin lesion classification remains challenging because of visual overlap and heterogeneous acquisition conditions. This study compared eight representation strategies for basal cell carcinoma, squamous cell carcinoma, cutaneous leishmaniasis, pressure ulcer, and venous ulcer. Global, attention-guided regional, fusion-based, and modern deep representations from DeiT-Small, EfficientNet-B0, EfficientNetV2-S, ConvNeXt-Tiny, and Swin-Tiny were evaluated with support vector machines under a group-aware protocol including duplicate auditing, repeated cross-validation, multiseed fine-tuning, source-aware analyses, and a held-out test. Swin-Tiny achieved the highest OOF Macro-F1 (0.9466), a repeated-cross-validation mean of 0.9463 ± 0.0015, and a test Macro-F1 of 0.9333, without significant superiority over ConvNeXt-Tiny. Regional and fusion strategies did not outperform the global representation. Source information remained strongly encoded, and SCC-to-BCC errors were concentrated in PAD-UFES-20. Thus, the results demonstrate strong internal discrimination but not source-invariant or external clinical generalization, supporting the need for independent multicenter validation.
No takes yet. Share an insight, caveat, or question.
Valles-Coral et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: