DVM-SLC: A Dual-View Meta-Aware Model for Reliable Multi-Class Skin Lesion Classification from Clinical and...

Background/Objectives: Skin lesion classification remains challenging in real clinical settings because diagnostically relevant information is distributed across different imaging modalities, class imbalance is often severe, and overall performance alone may conceal important weaknesses in model behavior. In this study, a dual-view meta-aware model, termed DVM-SLC was developed for multi-class skin lesion classification on the MILK10k dataset. Methods: The model jointly processes paired clinical and dermoscopic images and incorporates patient-level contextual information through a dedicated metadata branch. Cross-view information is integrated by a gated fusion mechanism to preserve modality-specific representations while allowing adaptive interaction between the two views. The model was evaluated not only in terms of comparative classification performance, but also with respect to calibration, robustness under common image corruptions, and expert-supported explainability. Results: Among the configurations evaluated using pooled out-of-fold predictions, the gated DVM-SLC achieved the highest Accuracy (0.7002) and the highest observed Macro-AUC (0.8774), whereas the gate-free dual-view model with metadata achieved the highest Macro F1-score (0.4174). The accuracy difference between the gated model and the dermoscopic-only baseline, which provided the highest non-proposed Accuracy, was statistically significant, whereas the observed Macro-AUC difference was not statistically significant. These findings indicated a metric-dependent trade-off rather than uniform superiority of a single fusion strategy. The raw confidence estimates were already reasonably aligned with empirical correctness, and temperature scaling provided only a modest, fold-dependent improvement in aggregate calibration. In robustness analysis, the strongest degradation was observed under Gaussian noise, while the model remained comparatively stable under blur, JPEG compression, brightness variation, contrast variation, and color shift. The explainability analysis was complemented by a structured dermatologist review. Gradient-weighted Class Activation Mapping (Grad-CAM) outputs from the clinical and dermoscopic branches were examined to determine whether the highlighted regions localized the lesion and corresponded to clinically meaningful morphological or dermoscopic features. This assessment was qualitative and was not intended as quantitative localization validation. Conclusions: Overall, the findings present DVM-SLC as a broadly evaluated approach with metric-dependent strengths and clear limitations, particularly in minority-class recognition. Its contribution lies in combining comparative performance analysis with calibration, robustness, and structured qualitative explainability assessment.