Uncertainty-Aware and Bias-Calibrated Vision Transformer–Graph Neural Network for Reliable Skin Cancer Detection

Abstract

Deep learning models for skin cancer diagnosis have improved, but most cannot quantify predictive uncertainty, are insensitive to dataset bias, and are unreliable under distributional shifts. Previous experiments using hybrid Vision Transformer–Graph Neural Networks (ViT–GNNs) with adaptive attention showed strong classification performance, but they were deterministic and made over–confident, biased predictions. We propose a hybrid ViT–GNN model with adaptive uncertainty and bias attention to improve skin cancer detection reliability and clinical acceptability. A supplementary prediction head estimates epistemic and aleatoric uncertainty using Monte Carlo dropout and deep ensembles, extending the Region-Adaptive Attention ViT–GNN architecture. A fairness calibration module uses group-wise temperature scaling and regularization to reduce performance disparities across skin tones and imaging situations. An out-of-distribution (OOD) detection system identifies anomalies and unfamiliar inputs to avoid overconfidence in wrong predictions. A clinical risk stratification layer classifies predictions as low, medium, or high risk based on confidence and uncertainty. The framework was evaluated on ISIC 2020, HAM10000, and PH2 datasets and cross-domain validated on unseen dermoscopic sources. The Expected Calibration Error, Brier score, fairness gap, and OOD detection metrics assess reliability, calibration, fairness, and safety beyond accuracy metrics. The empirical results suggest that the proposed model improves calibration and fairness without lowering competitive diagnostic accuracy, creating a reliable, interpretable, and ready-to-deploy skin cancer screening system.     Received: 14 May 2026 | Revised: 15 July 2026 | Accepted: 24 July 2026   Conflicts of Interest The authors declare that they have no conflicts of interest to this work.   Data Availability Statement The data that support the findings of this study are openly available in Kaggle at https://www.kaggle.com/datasets/nour12347653/skin-disease-detection-dataset-ham10000-isic.   Author Contribution Statement Aswani Dogga: Conceptualization, Methodology, Validation, Formal analysis, Data curation, Writing – original draft, Writing – review & editing, Visualization. Sivasubramanian R.: Software, Investigation, Resources, Supervision. Shanthi S.: Project administration.