Uncertainty-aware multimodal transformer with quantitative explainability for skin cancer diagnosis and risk stratification
Results in Engineering · 2026-12 · 32 · 113780
DOI: 10.1016/j.rineng.2026.113780 ↗Kurz gefasst
Fachartikel in der Zeitschrift Results in Engineering, 2026-12. Verfasst von Inzamam Mashood Nasir und 3 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Background
Reliable skin cancer diagnosis is difficult due to large variances in skin lesions, patient populations, and imaging conditions. The absence of sound uncertainty assessment and clinically relevant risk assessment of existing artificial intelligence (AI) applications presents further barriers to their clinical adoption.
Objective
This study aims to develop an integrated framework with an AI-based uncertainty and risk assessment module for skin cancer diagnosis and clinical risk assessment, using skin-lesion images and associated patient data.
Methods
This framework integrates a vision transformer (ViT)-based image representation, metadata encoding, cross-modal transformers, risk-aware prediction, and uncertainty assessment with Monte Carlo dropout, complemented by explainability assessment using Grad-CAM and the Percentage Attribution Index (PAI). A rigorous cross-dataset evaluation was conducted to benchmark the proposed module against the PAD-UFES-20 dataset and the independent MRA-MIDAS dataset, without any target-domain adaptation. Comprehensive evaluations include state-of-the-art performance comparisons, ablation analyses, uncertainty assessment, explainability assessment, and failure analyses.
Results
The proposed framework achieved an accuracy of 94.83%, an F1-score of 93.99%, a ROC-AUC of 97.21%, and a risk-stratification accuracy of 92.74% on the MRA-MIDAS dataset, demonstrating strong diagnostic and risk-prediction performance. The explainability assessment yielded a PAI score of 92.84%, indicating strong spatial correspondence between model attribution and lesion regions. The uncertainty analysis further demonstrated increased predictive uncertainty for ambiguous and challenging cases, providing complementary information regarding prediction reliability.
Conclusion
The results demonstrate the potential of uncertainty-aware multimodal learning for skin cancer classification and rule-based risk stratification under retrospective cross-dataset evaluation, while the uncertainty and explainability components provide complementary information about prediction reliability and attribution localization.
Zitieren
APA
Nasir, I. M., Tehsin, S., Damaševičius, R., & Ashraf, I. (2026). Uncertainty-aware multimodal transformer with quantitative explainability for skin cancer diagnosis and risk stratification. Results in Engineering, 32, 113780. https://doi.org/10.1016/j.rineng.2026.113780
Vancouver
Nasir IM, Tehsin S, Damaševičius R, Ashraf I. Uncertainty-aware multimodal transformer with quantitative explainability for skin cancer diagnosis and risk stratification. Results in Engineering. 2026;32:113780. doi:10.1016/j.rineng.2026.113780
BibTeX
@article{Nasir2026,
author = {Nasir, Inzamam Mashood and Tehsin, Sara and Damaševičius, Robertas and Ashraf, Imran},
title = {Uncertainty-aware multimodal transformer with quantitative explainability for skin cancer diagnosis and risk stratification},
journal = {Results in Engineering},
year = {2026},
volume = {32},
pages = {113780},
doi = {10.1016/j.rineng.2026.113780}
}
EUVIMED