S2200
Physics - Intra-fraction motion management and real-time adaptive radiotherapy
ESTRO 2026
Digital Poster 318
High-Accuracy Synthetic CT Generation from CBCT Using Deep Learning for Head and Neck Adaptive Radiotherapy Marcos Martinez Sanchez, Moises Saez Beltran, Carlos Ferrer Gracia, Concepcion Huertas Martinez, David García Riñon, Giorgia Yang Li, Miguel Barroso Miranda, Ivan Alonso Delgado, Pablo Galiano Fernandez, Pedro Jesus Perez Marcos, Rocio Gomez Robles, Daniel Rojo Navarrete, Raul Sanchez Lopez Medical Physics, Hospital Universitario La Paz, Madrid, Spain Purpose/Objective: Dose calculation for adaptive radiotherapy (ART) is limited on most linacs, as cone beam computed tomography (CBCT) images do not provide accurate Hounsfield units (HU) values for different tissues. Although new commercial ART solutions are available, their implementation remains limited, and new replanning CTs are often required in case of anatomical changes. This study aimed to develop a deep learning model capable of generating synthetic CT (sCT) from CBCT in head and neck patients, with sufficient HU accuracy for dose calculation, enabling its integration into ART workflows. Material/Methods: A total of 128 patients treated since 2023 were included. 90 were used for training and 38 for validation. For independent dosimetric evaluation in the planning system (TPS), and additonal cohort of 13 patients was selected. Imaging data consisted of the planning CT and the the first fraction CBCT, rigidly registered to the planning CT. Preprocessing included HU normalization and field-of-view (FOV) masks. The model was a U-Net3+-based GAN in a 2.5D configuration, with 11-slice stacks, edge maps, and FOV masks as inputs.Training employed a combination of perceptual losses (VGG, SSIM), geometric losses (edges, gradient), HU histogram consistency, region- specific losses (air, soft, bone, metal), and a Dice loss over the body contour. Validation included review of the sCT-CBCT fusion, image metrics (global and regional mean absolute error (MAE), body vs. background Dice, mean HU shift, histograms), TPS comparison of plans on the sCT versus the planning CT deformed to the CBCT, and 3D gamma analysis. Fig1. Example of sCT generation from CBCT input. Results: The model achieved an overall MAE of 18.4 HU, with mean errors of 7.9 HU in air, 22.9 HU in soft tissue, 27.0 HU in bone, and 67.4 HU in dense bone. The body to background Dice coefficient was 0.999 and the
The winning submission, Treat'n'Track, achieved mean and std DSC of 0.89±0.05, HD95 of 4.2±1.8mm, ECD of 1.5±0.99mm, TPF 42ms. Top-5 approaches reached DSC>0.87, HD95<5.64mm, and ECD<2.1mm, comparable to inter-observer variability (mean DSC 0.89, HD95 4.2mm, ECD 2.7mm). A common feature among leading submissions was fine-tuning of promptable foundation models, such as SAM2.For the Top-5, pelvic targets demonstrated higher fidelity to ground truth labels than thoracic and abdominal targets. Only one submission showed minor differences across B-field strengths. All submissions were amply fast for real-time application. Ranking proved robust to variations in dataset and metric selection. Conclusion: TrackRAD2025 demonstrated that foundation models fine-tuned on an MRI-linac-specific dataset achieve accurate 2D-CineMRI tumor tracking. Leading methods surpassed inter-observer variability at run-times amply sufficient for real-time tracking for common types of MRI-linacs. References: 1 Uijtewaal et al. "Dosimetric Evaluation of MRI - guided Multi - leaf Collimator Tracking and Trailing for Lung Stereotactic Body Radiation Therapy." (2021) Medical Physics. DOI: 10.1002/mp.147722 Keall et al. "AAPM Task Group 264: The safe clinical implementation of MLC tracking in radiotherapy." (2020) Medical Physics. DOI: 10.1002/mp.146253 Landry et al. "TrackRAD2025: Real-time tumor tracking for MRI-guided radiotherapy." (2025) DOI: 10.5281/zenodo.15044966. URL: https://trackrad2025.grand-challenge.org / https://trackrad.ch4 Wang et al. "TrackRAD2025 challenge dataset: real-time tumor tracking for MRI- guided radiotherapy." (2025) Medical Physics. DOI: 10.1002/mp.17964 Keywords: Real-time tumor localization
Made with FlippingBook - Share PDF online