ESTRO 2025 - Abstract Book

S3407

Physics - Machine learning models and clinical applications

ESTRO 2025

Oncologico (CRO) di Aviano of Florence, Aviano, Italy. 5 Medical Physics Unit, Azienda Ospedaliero-Universitaria Careggi, Florence, Italy

Purpose/Objective: The complexity of external beam photon radiotherapy has increased, raising the risk of errors and the need for advanced dose verification systems [1, 2]. Electronic Portal Imaging Devices (EPIDs), widely used for in-vivo dosimetry (IVD) and for pre-treatment verification, capture 2D dose information. 3D dose reconstruction is possible via physical models or Monte Carlo simulations [3, 4]. Artificial intelligence, in particular Deep Learning (DL), offers a promising alternative, matching Monte Carlo accuracy [5 – 7]. Existing studies are mostly based on simulated data. This work aims at developing a DL-based 3D in-vivo dose reconstruction framework using real EPID images acquired during treatment delivery. Material/Methods: A two-phase approach was adopted. First, a database of 229 measured EPID images and corresponding Portal Dose (PD) calculated by the TPS was created. A 2D U-net [8] was developed to map EPID images into PDs, serving as an alert system for discrepancies between predicted and reference PDs. Second, for 3D dose reconstruction, a framework including phantom CT volumes was implemented. A multi-input 3D CNN was designed to detect deviations between predicted and reconstructed doses, using a dataset of 70 phantom CTs, EPID images, and corresponding 3D dose distributions. Finally, the quality of the predicted doses was assessed via local gamma analysis [9].

Results: 2D framework achieved a predictive accuracy of 99% in dose distribution for the best cases, with a gamma-index criteria of 3mm/3%. The median gamma passing rate across all test cases (35 samples) was 98.41% demonstrating the model’s robustness across different scenarios. The dose predictions completed within 1 second, significantly faster than TPS calculations (≈ 20– 30 minutes). Preliminary 3D network prototypes are under evaluation, demonstrating promising results for multi-input dose reconstruction. These findings highlight the potential of DL for accurate and efficient dose prediction in radiotherapy, using EPID images and CT scans as input.