Automated multiclass skin lesion classification remains challenging because of severe class imbalance, substantial visual similarity between diagnostic categories, variable dermoscopic image quality, and overconfident predictions. This study presents DermaFuse-Lite, a classical hybrid framework that combines complementary deep and handcrafted decisions through quality- and uncertainty-guided adaptive fusion. The deep branch employs MobileNetV3-Large with a convolutional block attention module to capture discriminative spatial and channel-wise information. In parallel, the handcrafted branch uses 95 descriptors representing lesion geometry, asymmetry, colour distribution, local binary patterns, and grey-level cooccurrence statistics, followed by an XGBoost classifier. Five image-quality indicators and the predictive entropies of both branches are supplied to a learnable gate that assigns an image-specific fusion weight. A risk-aware objective discourages clinically consequential melanoma-to-nevus confusion, while temperature scaling improves confidence calibration. The framework was evaluated on all 10,015 HAM10000 images using five-fold stratified lesion-level cross-validation with group-disjoint inner validation. It achieved 83.75 ± 0.86% accuracy, 71.16 ± 3.95% balanced accuracy, a macro-F1 score of 0.721 ± 0.029, and a macro one-versus-rest AUC of 0.965 ± 0.003. Pooled out-offold analysis yielded a melanoma recall of 56.78% and reduced melanoma-to-nevus errors from 35.94% in the reference configuration to 32.17%. These findings support quality-aware adaptive fusion as a practical mechanism for integrating data-driven and interpretable representations while preserving calibrated, riskconscious predictions.

skin lesion,dermoscopy,Adaptive Fusion,image quality,predictive,AI model,AI