S2301
Physics - Machine learning and AI algorithms
ESTRO 2026
comparison as a potential biomarker for individualized risk assessment in breast radiotherapy. References: [1] A. Fodor et al., doi: 10.1016/j.clbc.2021.11.008, 2022.[2] M. M. Vincenzi et al., doi: 10.1016/j.radonc.2024.110700, 2025.[3] M. G. Ubeira- Gabellini et al. “Development and external multicentric validation of a deep learning-based clinical target volume segmentation model for whole-breast radiotherapy”, 2025.[4] M. J. Cardoso et al., doi: 10.48550/arXiv.2211.02701, 2022.[5] F. Pedregosa et al. “Scikit-learn: Machine Learning in Python”, 2011.[6] T. Chen and C. Guestrin. “XGBoost: A Scalable Tree Boosting System”, 2016.[7] T. Akiba et al., doi: 10.48550/arXiv.1907.10902, 2019.[8] A. Fodor et al., doi: 10.1016/j.breast.2020.12.004, 2021. Keywords: Machine Learning, Breast Cancer, Segmentation Digital Poster 2987 Training foundational vision models on 152,000 CBCTs with limited computing resources Alejandro Cortina Uribe 1 , Sidsel Poulsen 2 , Ivan Vogelius 2 , Jens Petersen 1 1 Computer Science, University of Copenhagen, Copenhagen, Denmark. 2 Oncology, Rigshospitalet, Copenhagen, Denmark Purpose/Objective: With self-supervised pre-training methods and large amounts of unannotated data, we can create foundational models that learn general-purpose features that are useful for relevant downstream tasks [1]. Often in RT departments, we have access to big uncurated datasets, but computational resources are limited. We proposed a simple strategy to predict model convergence in the pre-training stage, facilitating hyperparameter selection for limited computing resources, as well as tracking training efficiency. Finally, we validated the pre-trained model on a simple clinical downstream task in comparison to supervised training from scratch. Material/Methods: We extracted 152,000 3D CBCT scans from 7,365 patients treated with image-guided RT for lymphoma, head and neck, esophageal, and lung cancer at a single RT center, between 2009 and 2024. With this data, we trained a 3D vision transformer (ViT) image model using a masked image modeling task [2]. To select the model size, i.e. ViT-tiny (6.9M params) or ViT- small (24.4M params), we fitted a power-law curve to the validation loss of five 10-epoch repeated experiments. Once the model size is selected, we trained the model for 100 epochs, evaluating the quality of the learned representations by computing
tuning.Final models were evaluated using ROC-AUC, precision-recall curves, and SHAP-based feature explainability. Results:
As shown in Figure 1, for oedema prediction, the best- performing model (HistogramGradientBoost) achieved AUCs of 0.64 (train) and 0.56 (test), with mean cross- validation AUC of 0.57. DSC and VD were the most predictive features across all algorithms. SHAP analysis revealed that larger VDs (i.e.: clinical CTVs less extended than the reference) were associated with lower oedema risk, while higher DSC values correlated with increased oedema incidence.
As shown in Figure 2, for local relapse, the best- performing model (Logistic Regression) indicated a (non-significant) trend between smaller clinical CTVs relative to AI-predicted volumes and reduced risk. SHAP values confirmed a statistically significant contribution of cranial and caudal volume extensions to relapse probability, (AUC=0.69). Conclusion: Despite previously reported strong association with several clinical predictors[1,8], discrepancies between clinician-drawn and AI-predicted CTV contours are associated with post-radiotherapy outcomes. These findings support further validation of AI-based contour
Made with FlippingBook - Share PDF online