ESTRO 2025 - Abstract Book

S3445

Physics - Machine learning models and clinical applications

ESTRO 2025

4057

Digital Poster Quasi-3D virtual PSQA Using GAN-predicted 2D Dose Maps Samuele Cavinato 1 , Paola Ceroni 1 , Angelo Giannone 2 , Riccardo Lombardi 2 , Marta Paiusco 1 , Erica Simeone 2 , Nicola Zancopè 2 , Alessandro Scaggion 1 1 Medical Physics Department, Veneto Institute of Oncology IOV-IRCCS, Padova, Italy. 2 Department of Physics and Astronomy 'Galileo Galilei', University of Padova, Padova, Italy Purpose/Objective: We developed a virtual plan-specific quality assurance (vPSQA) method by training a Generative Adversarial Network (GAN) to predict 2D dose maps measured with a phantom array. This innovative approach replicates conventional measurement-based clinical workflows, preserving the full informational content of measurements while significantly reducing workload. Material/Methods: A dataset of 183 VMAT plans (60 head-and-neck SIB, 60 lung, 63 prostate) created in Eclipse v16.1 (Varian, Palo Alto, CA) and delivered on a TrueBeam STx system was collected. Quasi-3D plan-specific QA measurements were performed using the ArcCheck (SunNuclear, Melbourne, FL) dosimeter. The model, based on a pix2pix-GAN with a modified U-Net architecture featuring an encoder (downsampler) and a decoder (upsampler) [1], was trained on 126 plans and validated on 57. It predicts ArcCheck 2D dose maps by learning from pairs of ArcCheck-measured and TPS-calculated maps, discretized into 256 gray levels. Training ran for 25,000 steps. Model performance was evaluated by calculating local percentage dose differences between measured and GAN-predicted maps. Gamma passing rates (PRs) and mean gamma indexes (mGIs) were calculated between calculated and both measured and GAN-predicted maps to assess prediction accuracy. Results: Figure 1 compares TPS-calculated, measured, and GAN-predicted dose maps for a test case. Across the dataset, the average dose difference between GAN-predicted and measured maps was 1.9% (SD = 4.8%; median = 2.1%) for doses >10% of the maximum measured dose. The local percentage dose difference decreased with increasing dose values. These results are expected to improve with additional training and finer dose map discretization.