Trust-aware graph network shows promise and limits in skin lesion classification

Researchers have introduced TrustGraphNet, a graph neural network framework designed to make automated skin lesion classification more transparent and reliable, and their results reveal both genuine strengths and uncomfortable weaknesses. The framework, described in Discover Artificial Intelligence, integrates trust-weighted graph propagation, probabilistic uncertainty estimation, graph-based explanation signals, and a composite Trust Reliability Index into a single pipeline for dermoscopic image analysis. Rather than presenting only flattering metrics, the authors report in detail where the system underperformed a simple baseline, framing the work as a diagnostic evaluation of trust-aware learning rather than a claim of clinical readiness.

The motivation stems from a persistent gap in medical image analysis. Post-hoc explanation methods such as LIME, SHAP, and Grad-CAM are applied after a model has already produced its prediction, and systematic reviews conclude that the resulting explanations are frequently qualitative, inconsistently validated, and not demonstrably faithful to the underlying computation. Meanwhile, uncertainty quantification techniques such as Monte Carlo dropout and deep ensembles have developed largely as a separate research thread. A clinician using most current systems receives either a confidence score with no account of which evidence produced it, or an explanation heatmap with no account of whether that explanation itself should be trusted. TrustGraphNet attempts to connect these threads by making trust an explicit property of the learning process itself.

The architecture begins with adaptive trust-aware lesion graph construction. Each dermoscopic image is partitioned into superpixel regions, which become nodes in a graph. Three similarity channels capture spatial proximity, semantic cosine similarity of learned features, and morphological similarity based on area, perimeter, compactness, and border irregularity. A learned trust coefficient, estimated for every candidate edge before graph construction, is combined with these similarities through softmax-gated attention, and only the most reliable neighbors are retained. The authors argue this mirrors the inherently relational nature of dermoscopic diagnosis, in which evaluation of a lesion takes into account variability in dermal pigmentation, edge irregularities, the vascular network, and the pigmentation network in the context of surrounding skin.

During message propagation, the framework’s trust-guided relational propagation operator distinguishes relational importance from relational reliability. Attention coefficients determine which neighboring regions are informative, while trust coefficients regulate how strongly those regions contribute during aggregation, with trust refined layer by layer. Each node is additionally modeled as a probabilistic latent variable rather than a deterministic embedding, so uncertainty is represented within the graph reasoning itself. Node importance scores are discounted by uncertainty, meaning explanation confidence degrades on ambiguous lesions instead of producing a confidently wrong heatmap. A Trust Reliability Index then fuses confidence, graph uncertainty, explanation reliability, and structural consistency through learned softmax-gated weights, optimized jointly with the classification loss.

The experimental results are notably candid. On a corrected 203-image test set drawn from a 1,042-image HAM10000 cohort, TrustGraphNet achieved a three-seed mean accuracy of 0.608 plus or minus 0.010 versus 0.571 plus or minus 0.048 for a same-budget convolutional neural network. However, the framework had a lower mean macro-F1 of 0.194 versus 0.256, lower AUROC of 0.779 versus 0.802, and less favorable calibration. The authors emphasize that the three seeds share the same test set, so the spread reflects training variability rather than independent validation, and that no seed-level comparison survived multiple-testing correction. The work also uncovered a data integrity issue: an initial image-wise split had placed different images of the same physical lesion into both training and test sets for roughly 21 percent of test images, prompting a lesion-level resplit with zero cross-lesion leakage.

Perhaps the most striking negative finding concerns melanoma. At one seed, TrustGraphNet correctly identified none of the 38 melanoma cases in the HAM10000 test set, while the simple CNN baseline managed a recall of 0.18. Closer inspection showed the model did assign melanoma a higher probability to true melanoma cases than to benign ones, but the benign class dominated the final decision, suggesting a decision-boundary problem driven by class imbalance rather than an absence of learned signal. On the MILK10k subset of 2,501 images spanning 11 classes, the picture was worse still: TrustGraphNet’s melanoma recall remained 0.000 while the CNN achieved 0.488, and post-hoc logit adjustment offered little help there. The authors explicitly state that whatever mitigated the problem on HAM10000 did not transfer to MILK10k, and they do not claim to have identified the cause.

The composite Trust Reliability Index also failed its intended test. Historical means were 0.9547 for correct predictions and 0.9593 for incorrect predictions, meaning the index was slightly higher for wrong answers and saturated near 0.96 in both groups. The authors treat this as a negative result and withdraw any claim that the index discriminated errors. Because the original per-case outputs and the code implementing the four-component index were not recoverable, they declined to fabricate replacement statistics, an unusual and commendable stance in a field often criticized for selective reporting.

To address reviewer-requested analyses without misrepresenting unavailable historical outputs, the team 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 an AUROC of 0.767 at 64 by 64 resolution and 0.791 at 128 by 128, though the improvement over simple confidence-based ranking was not statistically established. Uncertainty-based selective ranking did behave as hoped: error rates fell from about 0.38 at full coverage to roughly 0.20 when the model abstained on the least confident half of cases. A validation-selected melanoma logit adjustment, chosen only on the 150-image validation split before any test evaluation, raised test melanoma recall from 0.385 to 0.692 at 64 by 64 while accuracy slipped only from 0.616 to 0.606, illustrating a classic sensitivity-specificity trade-off rather than a fix for the underlying imbalance.

The reproduction also tested whether the graph’s trust signals carry real explanatory weight. Using an equal-budget deletion procedure, the researchers removed the 20 percent of nodes with the highest trust scores and compared the resulting drop in prediction confidence against removing 20 percent of random nodes. High-trust deletion produced a mean probability decrease of 0.0508 versus 0.0055 for random deletion at 64 by 64, and 0.0565 versus 0.0013 at 128 by 128, indicating the model is measurably more sensitive to the regions its trust mechanism highlights than to arbitrary ones. The authors are careful to note this demonstrates perturbation sensitivity of the reproduced graph mechanism, not proof that the explanations are clinically meaningful. A resolution experiment further showed that increasing image resolution to 128 by 128 with roughly 50 nodes did not uniformly improve classification, raising accuracy only marginally while degrading macro-F1 and increasing computational cost.

The overall verdict the authors offer is deliberately measured. TrustGraphNet demonstrates that trust-aware graph learning is feasible and that its uncertainty signals can usefully rank cases for human review, but the evidence does not establish clinical reliability. Calibration was worse than a simple baseline, class sensitivity was erratic across seeds and datasets, the original trust index did not discriminate errors, and melanoma detection failed in one dataset entirely. The authors call for a redesigned, bounded trust index evaluated on untouched test sets, larger lesion-disjoint cohorts, additional seeds, stronger contemporary baselines, and prospective clinical data before any deployment-oriented claims can be made.

Subject of Research: Trust-aware graph learning with uncertainty estimation for dermoscopic skin lesion classification

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

Article References: Alwaeli, S. H., Jumaah, N. S., Jabbar, M. S., Khan, M. I., Bahaj, S. A. O., & Ayesha, N. (2026). Explainable trust aware graph learning for skin lesion classification with uncertainty estimation. Discover Artificial Intelligence, 6(1), Article 1644. https://doi.org/10.1007/s44163-026-02624-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02624-y

Keywords: TrustGraphNet, graph neural networks, skin lesion classification, dermoscopy, uncertainty estimation, explainable AI, melanoma detection, HAM10000, MILK10k, trust reliability index, calibration, medical imaging AI