S2042
Physics - Image acquisition and processing
ESTRO 2026
University Medical Center, Nijmegen, Netherlands. 9 Department of Radiation Oncology, University Medical Center Groningen, Groningen, Netherlands. 10 Department of Radiation Oncology, LMU University Hospital, LMU Munich, Munich, Germany. 11 Bavarian Cancer Research Center, (BZKF), Munich, Germany Purpose/Objective: Radiotherapy (RT) relies on accurate patient anatomy information for precise treatment planning. However, the lack of accurate electron density or stopping power in MRI and cone-beam CT (CBCT) hinders the clinical adoption of MRI-only workflows and MRI/CBCT- based daily adaptive RT. Synthetic CT (sCT) generation, particularly using deep learning, offers a promising solution by predicting CT HUs from MRI or CBCT [1]. The SynthRAD2025 Grand Challenge was organized to benchmark state-of-the-art sCT methods across diverse anatomical regions. Material/Methods: The SynthRAD2025 challenge (https://synthrad2025.grand-challenge.org/) invited participants to develop fully automated algorithms for two image synthesis tasks: Task 1 (MRI-to-sCT) and Task 2 (CBCT-to-sCT), addressing clinical needs in MRI- only and online adaptive RT. The dataset comprised multi-institutional paired images from head-and-neck, thorax, and abdomen cancer patients, totaling 2362 cases. Data was split into training (n=1533), validation (n=237), and test (n=592) sets, with a balanced representation from each anatomical site and institution. Participants trained their models on the publicly available dataset [2] and submitted their trained models to the challenge hosting platform. sCT accuracy was assessed using image similarity metrics: mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and multi-scale structural similarity index (MS- SSIM), anatomic fidelity of organs of interest: Dice similarity coefficient and Hausdorff distance (HD), and clinically relevant dose metrics: mean absolute dose errors, 3D gamma analysis (2%/2mm), and dose- volume histogram (DVH) based metrics. Teams were ranked for each metric, and a final rank was calculated by averaging the ranks across all metrics (Figure 1).
Results: The challenge attracted interest, with over 670 registrants and 29 valid submissions (14 for Task1, 15 for Task2). The winning MRI-to-sCT algorithm achieved an average MAE of 64.8 ± 21.2 HU and photon/proton gamma pass rates of 98.3±5.4% and 84.0±10.5%, respectively. For CBCT-to-sCT the top submission yielded an MAE of 48.2±13.3 HU and gamma pass rates of 99.3±1.2% (photon) and 88.6±7.9% (proton). Figure 2 presents sCTs of the top 3 teams of both tasks for an exemplary thorax and head-and-neck case. The top-performing methods revealed a prevalence of 3D encoder-decoder-style networks and 3D generative adversarial networks (GAN).
Made with FlippingBook - Share PDF online