S2292
Physics - Machine learning and AI algorithms
ESTRO 2026
Proffered Paper 2224
Evaluation of Uncertainty Quantification Methods for Deep Learning Normal Tissue Complication Probability Models in Head and Neck Cancer Patients Daniel C. MacRae 1 , Joëlle E. van Aalst 1 , Luuk van der Hoek 1 , Suzanne P.M. de Vette 1 , Hendrike Neh 1 , Matias A. Valdenegro Toro 2 , Johannes A. Langendijk 1 , Nanna M. Sijtsema 1 , Peter M.A. van Ooijen 1 , Lisanne V. van Dijk 1 1 Department of Radiation Oncology, University Medical Centre Groningen, Groningen, Netherlands. 2 Bernoulli Institute, University of Groningen, Groningen, Netherlands Purpose/Objective: Deep learning (DL) has recently emerged as a powerful tool for normal tissue complication probability (NTCP) modelling using full 3D dose distribution and imaging data. Despite their potential, the clinical adoption of DL NTCP models remains limited, as variability in patient-level accuracy raises concerns regarding the reliability and trustworthiness of their predictions. As shown in Figure 1, uncertainty quantification (UQ) could help address these issues by estimating prediction confidence through a distribution of predictions. Unlike auto-segmentation, where a physical ground truth exists and clinicians can verify and correct model outputs, NTCP modelling lacks a definitive ground truth for side effect outcomes. This absence of an objective reference underscores the need for reliable UQ to assess model confidence and support clinical interpretation. Existing applications of UQ in DL NTCP modelling remain limited and methodologically diverse [1]. To address this heterogeneity of methods, this study systematically evaluates multiple UQ approaches for DL NTCP models in HNC.
Conclusion: This multi-stage XGBoost workflow represents a significant methodological advancement. Its true power lies in its flexibility and generalizability. Having been validated for predicting complex clinical outcomes, we demonstrate its utility as a foundational platform for multi-omic predictive modeling. Combining this robust build process with SHAP, interpretability moves beyond single-task models, empowering clinicians with a transparent AI partner to enhance complex clinical decisions. References: [1] Bonci EA, Bandura A, Dooley A, et al. Artificial intelligence in NSCLC management for revolutionizing diagnosis, prognosis, and treatment optimization: A systematic review. Crit Rev Oncol Hematol. 2025 Sep 16;216[2] Chen, T. & Guestrin, C. XGBoost: A Scalable Tree Boosting System. in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 785–794[3] Lundberg, S. M., & Lee, S. I. (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems, 30[4] Lococo F, Boldrini L, Diepriye CD, et al. Lung cancer multi-omics digital human avatars for integrating precision medicine into clinical practice: the LANTERN study. BMC Cancer. 2023 Jun 13;23(1) Keywords: NSCLC, Interpretable AI, Multi-omics
Made with FlippingBook - Share PDF online