S2066
Physics - Image acquisition and processing
ESTRO 2026
Proffered Paper 2098
Improved synthetic CT generation using surface scan information integrated into a deformable registration algorithm for limited-Field of View data Joachim Marichal 1,2 , Ola Weistrand 3 , Geert De Kerf 1,2 , Michaël Claessens 2 , Stina Svensson 4 1 Faculty of Mecidine and Health Sciences, University of Antwerp, Antwerp, Belgium. 2 Department of Radiation Oncology, Iridium Network, Antwerp, Belgium. 3 RaySearch, RaySearch Laboratories AB, Stockholm, Sweden. 4 RaySearch, RaySearch, Stockholm, Sweden
Purpose/Objective: Cone Beam CT (CBCT)-based Adaptive
Radiotherapy requires reliable synthetic CT (sCT) images. Commercial sCT solutions exist for plan adaptation and dose recomputation [1],[2], but the CBCT’s limited Field of View (FoV) can cause inaccuracies when parts of the target volume fall outside the FoV, potentially with clinical relevance.While deep learning approaches have been explored to infer missing anatomy [3], limited work has investigated incorporating surface scan data when using deformable image registration (DIR) to reconstruct regions outside the FoV. This study evaluates the integration of Surface Guided Radiotherapy (SGRT) information to enhance sCT generation for breast cases with missing tissues on CBCT. Material/Methods: A retrospective analysis was performed on ten anonymized breast patients using a research version of RayStation v2025 (RaySearch). CBCT volumes were acquired on a Varian TrueBeam, with simultaneous surface scans from three room- mounted Catalyst (C-Rad) cameras.A full FoV CBCT (diameter 46 cm) was used to generate a reference synthetic CT (ground truth, GT). A smaller FoV CBCT (diameter 30 cm) was reconstructed from the same raw data to create two sCTs: a standard sCT (stan_sCT) and a surface-guided sCT (surf_sCT). The RaySearch virtual CT algorithm [1] was used for all sCTs. For surf_sCT, surface scan information was incorporated into the DIR to guide reconstruction in regions outside the FoV.The clinical target volume (CTV) was propagated from the planning CT to GT using DIR, then rigidly copied to stan_sCT and surf_sCT. Comparisons were made in terms of image similarity, external contour accuracy, and dose distribution. Results: surf_sCT showed superior agreement with GT for 13 of 15 evaluated metrics. As illustrated in Figure 1, surf_sCT exhibited reduced variability and a median gamma passing rate 2.5% higher than stan_sCT. Figure
Conclusion: The proposed model based on deep learning in this study can achieve precise sCT images from T1- weighted MRI images. The dosimetric differences between rCT images and sCT images are within the clinically acceptable criteria for stereotactic radiotherapy, and the visualization effect of the DRRs reconstructed based on sCT images in 6D-Skull tracking is comparable to that generated on rCT images, demonstrating the feasibility of a MR-Only workflow for the stereotactic radiotherapy of brain metastases on CyberKnife. References: [1] Li X, Jia L, Lin F, et al. Synthetic CT generation for pelvic cases based on deep learning in multi-center datasets[J].Radiation Oncology, 2024.DOI:10.1186/s13014-024-02467-w.[2] Miyato T , Kataoka T , Koyama M ,et al.Spectral Normalization for Generative Adversarial Networks[J].International Conference on Learning Representations, 2018.DOI:10.48550/arXiv.1802.05957.[3] Aljaafari L, Speight R, Buckley DL, et al.Clinical validation of using a commercial synthetic-computed tomography solution for brain MRI-only radiotherapy treatment planning[J]. Phys Imaging Radiat Oncol, 2025.DOI: 10.1016/j.tipsro.2025.100328 Keywords: Synthetic CT; CyberKnife; brain metastases;
Made with FlippingBook - Share PDF online