ESTRO 2025 - Abstract Book

S3385

Physics - Machine learning models and clinical applications

ESTRO 2025

1792

Digital Poster Deep learning decision-making framework for optimal technique selection in breast radiotherapy: predicting dose distributions for IMRT and 3D-CRT Jaime Perez-Alija 1 , Pedro Gallego 1 , Eva Ambroa 2 , Núria Jornet 1 , Cristina Ansón 1 , Natalia Tejedor 1 , Helena Vivancos 1 , Agustí Ruiz 1 , Marta Barceló 1 , Alejandro Domínguez 1 , Victor Riu 1 , Javier Roda 1 , Pablo Carrasco 1 , Simone Balocco 3,4 , Oliver Díaz 3,4 1 Radiofísica i Radioprotecció, Hospital de la Santa Creu i Sant Pau, Barcelone, Spain. 2 Medical Physics Unit, Radiation Oncology Department, Consorci Sanitari de Terrassa, Barcelone, Spain. 3 Departament de Matemàtiques i Informàtica, Universitat de Barcelona, Barcelone, Spain. 4 Computer Vision Center, Univeristat Autònoma de Barcelona, Bellaterra, Spain Purpose/Objective: This study aimed to develop a decision-making framework for breast radiotherapy using deep learning to predict dose distributions for intensity-modulated radiation therapy (IMRT) and three-dimensional conformal radiation therapy (3D-CRT), optimizing the treatment selection technique based on patient-specific anatomy and on predefined clinical criteria. Material/Methods: A 2D U-Net convolutional neural network was trained on a retrospective dataset, comprising 346 patients treated with either 3D-CRT or IMRT. Patient inclusion criteria encompass left and right breast with or without axillary and supraclavicular lymph nodes. The model predicted dose distribution maps. Dose-volume histogram (DVH) metrics for critical organs were created from the predicted dose maps. To validate the proposed framework, an independent medical physicist evaluated the plans and selected the optimal one from an external cohort of 30 patients. Both IMRT and 3D-CRT plans were manually created, and the physicist selected the optimal plan based on predefined criteria (V95% and mean dose for the PTVs; V25 Gy and mean dose for the Heart, and V20 Gy and mean dose for the ipsilateral Lung). Confusion matrices (CM) were generated to compare the framework’s recommendations with both historical clinical decisions and the independent observer’s choices. Results: The model demonstrated high concordance with clinical dose distributions, especially in organs at risk (OARs). Figure 1 presents an example of the predicted dose distribution for random patients, and the DVH comparisons. Table 1 contains two CM: the first one depicts a CM between the decisions made by the medical physicist (ground truth) against the clinical decisions made historically; the second one, shows a CM between the decisions made by the independent observer against the proposed decision-making framework. When compared to independent evaluations, the framework achieved accuracy, recall, and precision rates of 90%, 95.7%, and 91.7%, respectively, outperforming the clinical decisions actually made for those patients (accuracy 50%, recall 47.8%, precision 78.6%). The framework reliably predicted dose distributions for IMRT and 3D-CRT, enhancing the accuracy and consistency of treatment planning.