Accurate multi-class skin cancer classification is clinically valuable only when models perform reliably across both common and rare diagnostic categories. To address persistent long-tail failures, we systematically evaluate three enhanced vision-transformer architectures, GC-ViT Small, CoAt-Lite, and FocalNet, augmented with two novel attention modules: Transformer Blocks for global context integration and Separable Self-Attention for precision–recall calibration. Models were evaluated on held-out test sets from HAM10000 and ISIC 2019 under a revised 70/15/15 train/validation/test protocol. GC-ViT-E, which implements explicit global-context modeling, achieved 92.81% accuracy on HAM10000 and 91.42% accuracy on ISIC 2019, consistently outperforming hierarchical alternatives by approximately 2.5–4.0 percentage points in accuracy and 3.5–6.4 percentage points in macro-F1 while maintaining similarly high macro-AUC values (0.977–0.991). Crucially, GC-ViT-E preserved strong class-wise performance across all categories, including underrepresented lesions, with particularly high F1-scores for dermatofibroma and vascular lesions, while the most challenging classes remained actinic keratosis, dermatofibroma, and squamous cell carcinoma on ISIC 2019. Controlled ablation studies attribute the larger recall gains primarily to Transformer Blocks and show that Separable Self-Attention mainly refines the precision–recall balance with smaller effects on overall macro-F1. Cross-dataset experiments revealed pronounced directional asymmetry: training on ISIC 2019 and testing on HAM10000 yielded 97.54% accuracy and 0.9348 macro-F1, whereas the reverse direction dropped to 77.49% accuracy and 0.6633 macro-F1, underscoring the importance of source-domain diversity for external generalization. These results indicate that explicit global attention and attention factorization materially improve rare-class recognition, balanced multi-class decision quality, and transfer robustness in dermoscopic classification.
Alhaj et al. (Sat,) studied this question.