ESTRO 2025 - Abstract Book

S3408

Physics - Machine learning models and clinical applications

ESTRO 2025

Conclusion: This study explores the potential of using deep learning to convert EPID images into dose distributions for in-vivo treatment verification. The 2D model has shown very good accuracy and speed, while 3D models are in the training optimization phase. Starting from the results of the 2D model, we used the same dataset to extend the framework to 3D dose distribution prediction, integrating phantom CTs. Early experiments with 3D U-Net architectures are being expanded to advanced models like Vision Transformers. This work presents 2D model results and preliminary 3D findings, based on data from Careggi University Hospital of Florence, Italy.

Keywords: Radiotherapy, In-vivo dosimetry, Deep learning

References: 1) Koka et al, Cancer Manag Res , 2022 2) Shafiq et al, Radiother and Oncol , 2009 3) Wouter et al, Radiother Oncol , 2008 4) Olaciregui-Ruiz et al. Phys Imaging Radiat Oncol, 2020 5) Zhang et al, Radiat Oncol . 2022 6) Chan Maria et al, Front Artifi Intell , 2020 7) Li et al, Phys Med Biol . 2021

8) Ronneberger et al, 2015 9) Li et al, Med. Phys . 2011