Hypergraph-Driven Robust Vision Framework with Calibrated Explainability for Reliable Skin Lesion Analysis

Reliable automated skin lesion diagnosis remains challenging because deep learning systems often suffer from adversarial sensitivity, feature redundancy, poor calibration, and biased learning under imbalanced clinical data. This work introduces a unified framework named RA-HYPER-VISIONNet that combines hypergraph-guided representation learning with uncertainty-aware optimization to improve robustness, calibration, and interpretability in dermoscopic image analysis. The proposed architecture integrates adversarial distribution alignment, intrinsic explainability-guided feature refinement, Bayesian focal optimization, and transformer-based uncertainty fusion within a single learning pipeline. Unlike conventional post-hoc explanation methods, the framework embeds explanation consistency directly into feature learning to produce clinically meaningful predictions. Experiments were conducted on the ISIC2019, HAM10000, and ISIC2018 benchmark datasets for classification and segmentation tasks. The model achieved 97.56% classification accuracy with 0.991 AUC on ISIC2019 and 96.94% accuracy with 0.988 AUC on HAM10000. For lesion segmentation, a Dice coefficient of 94.96% was obtained on ISIC2018. Under adversarial perturbations, the framework preserved stable predictive performance with limited degradation. In addition, feature compression significantly reduced computational overhead while maintaining diagnostic precision. The results demonstrate that integrating robustness, calibration, and explainability into a unified architecture can improve the dependability of AI-assisted dermatological diagnosis.