Explainable trust aware graph learning for skin lesion classification with uncertainty estimation

TrustGraphNet integrates trust-weighted graph propagation, probabilistic uncertainty estimation, graph-based explanation signals, and a composite Trust Reliability Index (TRI) for dermoscopic skin-lesion classification. We evaluate the framework on executed HAM10000 (1,042 images; seven classes) and MILK10k (2,501 images; 11 classes) subsets using lesion-level separation where metadata permit. On the corrected 203-image HAM10000 test set, TrustGraphNet achieved a three-seed mean accuracy of 0.608 ± 0.010 versus 0.571 ± 0.048 for a same-budget CNN, but had lower mean macro-F1 (0.194 ± 0.073 vs. 0.256 ± 0.031), AUROC (0.779 ± 0.016 vs. 0.802 ± 0.014), and less favorable calibration. The historical TRI means were 0.9547 for correct and 0.9593 for incorrect predictions and therefore do not establish error discrimination. To address reviewer-requested analyses without misrepresenting unavailable historical per-case outputs, we additionally performed an independently trained lesion-disjoint reproduction on the verified 1,042-image/857-lesion cohort using a frozen 689/150/203 split. In that reproduction, uncertainty detected prediction errors with AUROC 0.767 (95% bootstrap CI 0.702–0.830) at 64 × 64 and 0.791 (0.725–0.852) at 128 × 128; validation-selected melanoma logit adjustment increased test melanoma recall from 0.385 to 0.692 at 64 × 64, with accuracy changing from 0.616 to 0.606; and equal-budget deletion produced a larger probability decrease for high-trust than random nodes (0.0508 vs. 0.0055). Increasing resolution to 128 × 128 did not uniformly improve classification. These findings support a diagnostic evaluation of trust-aware graph learning and identify limitations in the original TRI, calibration, class sensitivity, and reproducibility rather than establishing clinical reliability.