ESTRO 2025 - Abstract Book

S3784

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

[2] Klein S, Staring M, Murphy K, Viergever MA, Pluim JPW (2010). Elastix: a toolbox for intensity based medical image registration, IEEE Transactions on Medical Imaging, 29(1), 196-205 [3] Koopmans PJ, Monshouwer R, Findhammer J, Smeenk RJ, Verheij M, van der Bijl E. (2024) Systematic trends in T2w and ADC in hypofractionated prostate cancer treatments on a 1.5T MR-Linac. Proceedings of ESTRO-2024, p2033.

2464

Proffered Paper Deep learning-based prediction of radiation pneumonitis in advanced stage lung cancer using average 4D CT, 3D dose, and lung segmentations Robert van der Wal, Lous R. de la O Arevalo, Luuk van der Hoek, Suzanne P.M. de Vette, Hendrike Neh, Daniel C. MacRae, Johannes A. Langendijk, Nanna M. Sijtsema, Peter van Ooijen, Robin Wijsman, Lisanne V. van Dijk Radiotherapy, University Medical Center Groningen, Groningen, Netherlands Purpose/Objective: Radiation pneumonitis (RP) is a severe and potentially life-threatening toxicity following radiotherapy for lung cancer. Normal tissue complication probability (NTCP) models for RP, such as developed by Niezink et al. [1], are designed to estimate risk of toxicity, and may aid in clinical decision making for treatment selection and dose optimization during radiotherapy treatment planning. NTCP models rely on patient characteristics (e.g. gender, smoker) and discrete dose parameters of lungs excluding gross tumor volume (GTV) structure, thus they lack the ability to analyse 3D spatial information. This study developed a convolutional neural network (CNN) model incorporating 3D imaging and 3D-dose data to improve the prediction of RP in advanced stage lung cancer. Material/Methods: This study included 1031 advanced-staged lung cancer patients who received radiotherapy between 2013 and 2024 (Table 1). All patients were enrolled in a prospective, standardized follow-up program. The endpoint was defined as RP grade ≥2 (CTCAE v4.0) within one year after radiotherapy . The patient cohort was randomly divided into a development (70%) and independent test (30%), stratified by endpoint and adjuvant immunotherapy status. A deep learning (DL) model using prescribed dose distributions, average 4D planning CT, lungs-GTV segmentation and clinical variables was developed. The ResNet model was selected, trained and evaluated using 4-fold cross validation. Performance of the DL model was compared with the refitted NTCP model using Area Under the Curve (AUC). Furthermore, attention maps are investigated for interpretability of the DL model. Results: The incidence of physician-rated RP in the entire cohort was 14.9%. Hyperparameter tuning identified that a combination of the Adam optimizer, batch size 16 and binary cross-entropy (BCE) as loss function resulted in the highest AUC. The DL model demonstrated an AUC of 0.75 [95%CI:0.68-0.82] (Table 1). This performance was higher than the AUC achieved by the refitted NTCP model [0.73, 95% CI:0.68–0.81]. The DL model attention maps reveal larger regions of high attention in the lungs for high-risk RP patients. An overall trend was seen for more attention in the ipsi-lateral lung outside the high-tumor dose region (Figure 1). Further analyses are needed for more definitive conclusions.