ESTRO 2025 - Abstract Book

S3780

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

References: [1] van Amsterdam WAC, Elias S, Ranganath R. Causal Inference in Oncology: Why, What, How and When. Clin Oncol (R Coll Radiol). 2024 Jul 11:S0936-6555(24)00286-3. doi: 10.1016/j.clon.2024.07.002. [2] Textor J, van der Zander B, Gilthorpe MS, Liskiewicz M, Ellison GT. Robust causal inference using directed acyclic graphs: the R package 'dagitty'. Int J Epidemiol. 2016 Dec 1;45(6):1887-1894. doi: 10.1093/ije/dyw341.

[3] UK-based DAG: https://dagitty.net/mJjmwwCrd [4] US-based DAG: https://dagitty.net/mFhajxRT4

2377

Poster Discussion Transformer-based NTCP model for late taste loss Hendrike Neh, Suzanne P.M. de Vette, Luuk van der Hoek, Daniel C. MacRae, Peter M.A. van Ooijen, Nanna M. Sijtsema, Johannes A. Langendijk, Lisanne V. van Dijk Radiotherapy, UMCG, Groningen, Netherlands Purpose/Objective: Taste loss is a common and complex side effect of radiotherapy affecting over 25% of head and neck cancer (HNC) patients six months post-treatment. Accurate risk predictions from normal tissue complication probability (NTCP) models can guide treatment planning to minimize side effects, like taste loss, that significantly affect quality of life for HNC survivors. Advances in artificial intelligence (AI) have enabled the development of deep learning NTCP models that predict radiation-induced toxicity using comprehensive 3D imaging and radiation dose information. Convolutional neural networks (CNNs), commonly used for 3D inputs, are limited to local feature extraction, while emerging transformer-based models can learn global features of a full image and improve contextual understanding. This study aims to improve NTCP model performance for predicting late taste loss by implementing various CNN and transformer-based models, using the complete 3D dose distribution, CT scans, and organ-at-risk (OAR) segmentations. Material/Methods: The patient cohort included 972 HNC patients (2007 – 2022) with the endpoint defined as moderate-to-severe patient-rated taste loss at 6 months, as assessed on the EORTC QLQ-H&N35 4-point Likert scale (none, mild, moderate and severe) (Table 1). The cohort was split into a development (80%) and unseen test set (20%). The input for the deep learning model consisted of the 3D dose distribution, planning CT, and segmentations of all common head and neck OARs. Resnet and Vision Transformer (ViT) networks were trained using 5-fold cross-validation performed to ensure robustness. A previously published conventional late taste loss NTCP model with the predictors oral cavity mean dose, parotid gland mean dose, and age served as reference model [1]. Results: The ViT deep learning model showed better generalizability with area under the curve (AUC) performances on the independent test set of 0.70 (95%CI: 0.62-0.77) compared to the Resnet with an AUC of 0.67 (95%CI: 0.60-0.75). The refit of the reference model achieved an AUC of 0.70 (95%CI: 0.64-0.78) on the independent test set. The calibration plots showed better calibration for the reference model than the Resnet and ViT models (Figure 1).