ESTRO 2025 - Abstract Book

S3787

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

Material/Methods: An international cohort of HNC patients that received definitive (chemo)radiotherapy from two centres were used. 1,090 patients from a Dutch centre were used to develop (80%) and test (20%) the DL models, while 359 patients from a US-centre were used to externally validate the models. Five different toxicity endpoints were modelled: aspiration, dysphagia, sticky saliva, taste loss, and xerostomia, assessed at 6 months post-radiation therapy. All endpoints were patient-rated (Dutch-centre: EORTC QLQ-C30 & QLQ-HN35, US-centre: MDASI-HN), except for dysphagia, which was physician-rated (Dutch-centre: CTCAEv4.0, US-centre: PSS-HN diet normalcy). A novel DL network (TransRP) combining a DenseNet model and a transformer architecture was employed to train two types of NTCP models: (i) five independent single-tox models, and (ii) one multi-tox model predicting all five toxicities simultaneously (Figure 2). The multi-modal TransRP models used 3-dimentional distributions, CT scans, and organ-at-risk segmentations as image-based inputs, alongside a set of clinical predictors. Results: The multi-tox models demonstrated better performance overall than the single-tox models on the independent test set across all five toxicities (mean AUC test 0.74 vs. 0.72), yet this improvement was largely attributed to a large prediction improvement for aspiration (multi-tox model AUC test 0.76; single-tox model AUC test 0.62). The multi-tox model performed slightly worse than the single-tox model on sticky saliva (AUC test 0.66 vs. 0.68) and taste loss (AUC test 0.66 vs. 0.68). Both sets of DL models obtained lower performance on the external validation set (single-tox model AUC external 0.64, MT model AUC external 0.63) compared to the independent test set.

Conclusion: Although the multi-tox approach demonstrated certain improvements, it did not universally outperform the single tox models across all toxicities. Nevertheless, the proposed multi-toxicity modelling approach holds promise for advancing NTCP modelling by integrating inter-toxicity interactions, warranting further exploration.

Keywords: NTCP, head and neck, multi-label