S2285
Physics - Machine learning and AI algorithms
ESTRO 2026
radiotherapy-induced xerostomia in head and neck cancer patients using day-to-day kinetics of radiomics features. Phys Imaging Radiat Oncol. 2022;24:95- 101.5. Pan XB, et al. Prognostic nomogram of xerostomia for patients with nasopharyngeal carcinoma after intensity-modulated radiotherapy. Aging. 2020;12(2):1857-66. Keywords: Deep learning, predictive modelling
was trained and evaluated using k-stratified 5-fold cross-validation. Comparison metrics calculated were precision, recall, F1 score, accuracy, Area under the curve (AUC), Brier score, extended calibration error (ECE), and specificity.
Digital Poster 1316
Robust to Bladder Filling Variations: Deep learning–based DIR auto segmentation in MR guided online adaptive radiotherapy for prostate cancer patients Kota Abe 1 , Katsuhiro Hirano 2 , Masato Tsuneda 1 , Makoto Saito 3 , Yukio Fujita 1,4 , Yasukuni Mori 5 , Hiroki Suyari 6 , Takashi Uno 1 1 Graduate School of Medicine, Chiba University, Chiba, Japan. 2 Graduate School of Science and Engineering, Chiba University, Chiba, Japan. 3 Department of Radiology, Chiba University Hospital, Chiba, Japan. 4 Department of Radiological Sciences, Komazawa University, Tokyo, Japan. 5 Institute for Excellence in Educational Innovation, Chiba University, Chiba, Japan. 6 Graduate School of Engineering, Chiba University, Chiba, Japan
Results: Of 418 eligible patients, 316 had complete dosimetric data, of which 114 xerostomia cases were present. Table 1 shows xerostomia prediction model performance. TabPFN and RF demonstrated the highest AUC value (0.71). RF or TabPFN demonstrated best values for all metrics. Feature importance differed by model: contralateral submandibular and parotid gland doses dominated MVLR/RF, while pharyngeal constrictor doses were most important for TabPFN.
Purpose/Objective: In MR-guided online adaptive radiotherapy,
deformable image registration (DIR) is an essential component of the workflow for contour propagation. However, in prostate cancer patients, bladder volume variation often degrades the accuracy of DIR-based contour transfer. This study aimed to develop a deep learning–based DIR method that can adapt to bladder deformation and to evaluate its performance in clinical MR images. Material/Methods: A retrospective analysis was conducted on forty-one patients with intermediate risk prostate cancer who underwent five-fraction stereotactic body radiotherapy using the 1.5T MR-Linac system. The clinical target volume included prostate and base of SV 1 cm. Each patient underwent five treatment sessions; the planning MR was paired with each session’s daily MR, yielding five image pairs per patient and 205 pairs in total. Of these, 32, 5, and 4 patients were allocated to the training, validation, and testing cohorts, respectively. The proposed approach employed a TransMorph-based architecture, augmented with a voxel-wise weighted loss function that adaptively modulates local regularization strength to account for bladder deformation. The trained network estimated deformation vector fields between image pairs, which were subsequently used to propagate contours from
Conclusion: Models demonstrated AUC values within the published range (4,5) for ML models predicting xerostomia from clinical variables. RF and TabPFN models outperformed MVLR for all metrics suggesting better ability to handle additional features. Despite no feature selection or tuning, TabPFN achieved the joint- highest AUC value, supporting its utility as a robust, efficient classifier for NTCP prediction and potential for complex oncological datasets analysis. References: 1. Stieb S, et al. NTCP Modeling of Late Effects for Head and Neck Cancer: A Systematic Review. Int J Part Ther. 2021;8(1):95-107.2. Young T, et al. RT-HaND_C: A Multi-Source, Validated Real-world Head and Neck Cancer Dataset for Research. Clin Oncol (R Coll Radiol). 2025 Nov;47:103935.3. Kim JH. Multicollinearity and misleading statistical results. Korean J Anesthesiol. 2019;72(6):558-69.4. Berger T, et al. Predicting
Made with FlippingBook - Share PDF online