S2065
Physics - Image acquisition and processing
ESTRO 2026
Shanghai United Imaging Healthcare Co., Ltd, Shanghai, China
Purpose/Objective: To generate synthetic CT (sCT) from T1-MRI with a deep-learning model and validate an MR-only workflow for CyberKnife stereotactic radiotherapy (SRT) of brain metastases. Material/Methods: A total of 63 brain metastasis patients who had previously undergone CyberKnife SRT were included. Their T1-weighted MRI images and CT images were collected, with 50 for training and 13 for testing.We propose a Consistency Regularization Generative Adversarial Network (CRGAN) to improve the generalization of MRI-to-CT synthesis. The framework employs dual generators and discriminators, with a transformer integrated into the generator. An enhanced contrastive learning loss, incorporating semantic context, is introduced to strengthen structural consistency between MR and sCT. Furthermore, consistency regularization derived from semi-supervised learning is leveraged to boost model robustness and adaptability to heterogeneous data sources. Then treatment planning was redesigned based on sCT images. The differences in dosimetric parameters between the sCT-based plans and real CT- based plans were compared and analyzed. Finally, the Digitally Reconstructed Radiographs (DRRs) generated from real CT images (rDRRs) and those from sCT images (sDRRs) were each registered with the digital radiography images (DRs) captured during treatment. The translational differences were compared to assess the consistency of their positioning accuracy. Results: The model constructed in this study exhibited excellent performance in soft tissue contrast and bone regions. The average values of MAE and RMSE were 10.05 HU and 75.89 HU, respectively, while the average SSIM and PSNR were 0.98 and 34.75 dB, respectively. The average DSC values for key organ contours (brainstem, eyeballs, and optic nerves) were 88.14%, 90.49%, and 55.26%, respectively. Dosimetric analysis showed that the differences in PTV’s V100%, D95%, Dmax, and Dmean between the sCT-based plans and real CT-based plans were all less than 1.5%, and the dose deviations of OARs were within clinically acceptable ranges. The average translational differences in the X, Y, and Z axes between the registered rDRRs-DRs and sDRRs-DRs were 0.001 mm, 0.157 mm, and 0.012 mm, respectively.
Conclusion: This proof-of-concept single-detector cone-beam ion- CT system achieves improved acquisition times (15- 30s per-projection scan time) while maintaining sub- millimeter WEPL accuracy across three ion modalities. Helium ions demonstrated optimal performance with 0.02cm maximum WEPL deviations from ground truth .The simplified architecture offers a cost-effective opportunity toward clinical translation of ion-CT for potential application for treatment planning. References: [1] Paganetti H. Phys Med Biol 2012;57:R99-R117. [2] Volz L et al. Phys Med Biol 2021;66:245013. [3] Sadrozinski HFW et al. Nucl Instrum Methods A 2013;699:205-210. [4] Schneider T et al. Front Phys 2021;9:595721. [5] Wickert R et al. Cancers 2022;14:5865. Keywords: Ion CT, Helium, Proton
Digital Poster 1800
Feasibility Study of an MR-Only Workflow for CyberKnife Stereotactic Radiosurgery in Brain Metastases yajuan Wang 1 , Ziquan Wei 2 , Lecheng Jia 2 , Xue Bai 1 1 Radiation Physics, Zhejiang Cancer Hospital, Hangzhou, China. 2 Radiotherapy Business Unit,
Made with FlippingBook - Share PDF online