ESTRO 2025 - Abstract Book

S3442

Physics - Machine learning models and clinical applications

ESTRO 2025

References: [1] Oosterhof, F., Marmitt, G. G., Langendijk, J., & Both, S. (2024). TOWARDS AUTOMATED DOSE-GUIDED PATIENT TREATMENT ALIGNMENT QUALITY ASSURANCE AT CLINICAL TIMESCALE. International Journal of Particle Therapy , 12 , 100142. [2] H. Lee, et al., moquimc, https://github.com/mghro/moquimc (2024).

3982

Digital Poster Synthetic image prediction for patient-specific quality assurance in stereotactic radiosurgery of multiple brain metastases Nicola Lambri 1,2 , Martina Viganò 3 , Daniele Loiacono 3 , Ciro Franzese 1,2 , Marta Scorsetti 1,2 , Pietro Mancosu 1,2 1 Radiotherapy and Radiosurgery Department, IRCCS Humanitas Research Hospital, Rozzano, Italy. 2 Department of Biomedical Sciences, Humanitas University, Pieve Emanuele, Italy. 3 Dipartimento di Elettronica, Informazione e Bioingegneria, Politecnico di Milano, Milano, Italy Purpose/Objective: Stereotactic radiosurgery (SRS) for treating multiple brain metastases can be delivered using a single isocenter with multiple non-coplanar arcs. This approach, allowing for highly conformal dose distributions, requires extensive modulation of treatment machine parameters. Thus, SRS arcs are at higher risk of PSQA failure than standard VMAT treatments, potentially delaying treatment delivery. In this work, we developed a framework able to generate synthetic EPID fluence images from TPS ones, to predict the PSQA outcome (gamma passing rate – GPR) for SRS arcs. Material/Methods: The dataset comprised EPID-TPS fluence image pairs from 1593 radiation fields, optimized using HyperArc (Varian) and gathered from 358 SRS patients treated between Dec-2020 and Aug-2023 at our Institute. The model consisted of a conditional generative adversarial network (cGAN), with a U-Net like generator to output synthetic EPID images and a PatchGAN discriminator. Both components were conditioned on TPS fluence images. To enhance model performance, a second channel was added to highlight the orientation of the MLC and transmitted fluence. A bounding box detection algorithm was used to select pixels within each treatment beam eye view, and data augmentation (flip/rotation) was applied to images with large bounding box size, corresponding to more complex shapes. Data were divided into 1748 (473 augmented), 159, and 159 image pairs for training, validation, and testing. Model evaluation was conducted on the test set by comparing: (i) root mean squared error (RMSE) and peak signal to-noise ratio (PSNR) between real and synthetic EPID images; (ii) mean absolute error (MAE) between real and synthetic GPR using a 3%/1 mm criterion. Results: Synthetic EPID fluence images were visually indistinguishable from real EPID ones, with an RMSE of 0.02±0.01 CU and PSNR of 51±4. The median synthetic and measured GPRs were 99.2% and 98.4%, respectively, with half of the values falling respectively between 98.1-99.7% and 94.8-99.7%. A worse correspondence between real and synthetic gamma analyses was observed for treatment fields with high geometric complexity (Figure 1). Accordingly, the model performance deteriorated for treatment fields having measured GPR<90%, where the median synthetic GPR was 97.4% compared with a median measured GPR of 87.7% (Figure 2).