2192 - End-to-End Clinical Evaluation Testing of Synthetic CT for MRI-Only Brain and Pelvis Radiotherapy

Purpose/Objective(s): AI-generated synthetic CT (sCT) is an emerging reference imageset for MRI-only radiotherapy workflows, with potential to eliminate conventional CT imaging while maintaining dosimetric accuracy. This study aimed to validate a commercial AI-based sCT model for brain and pelvis MR-only radiotherapy by evaluating dosimetric performance, registration accuracy, and end-to-end clinical workflow verification.

Materials/Methods: A retrospective study was performed using brain and prostate patients previously treated at our Institution. sCT images were generated using the Siemens RT Image Suite VB60 deep-learning algorithm from sagittal (brain) and axial (pelvis) 3D T1w Dixon MRI sequences acquired on a Siemens 3T MRI. Treatment plans were generated using conventional CT images using Monaco (Elekta) and transferred to Mosaiq (Elekta) for delivery on an Elekta VERSA HD linac. The plans were then recalculated on sCT images using SciMoCa™ (Radiologica), and 3D gamma analysis was performed to compare CT and sCT dose distributions using x%/y mm (x=1,2,3; y=2,3) criteria. Dose differences between sCT and CT-based calculations were assessed. DRRs were generated from both sCT and CT and compared using structural similarity index map (SSIM). As part of end-to-end verification, new treatment plans were created on sCT images, promoted into Mosaiq, and delivered. CBCT imaging was performed, and image registration was conducted on the onboard XVI system using a phantom placed on the couch with sCT set the reference image.

Results: The dosimetric analysis demonstrated 3D gamma passing rates for brain and pelvis cases across different gamma passing criteria, as presented in the table. The overall dose differences between sCT and CT were ±2% for the brain and ±1% for pelvis cases. All sCT-based treatment plans were successfully promoted into Mosaiq. Additionally, sCT were successfully imported as reference images for position verification on XVI, and used to register CBCT images. DRR comparisons demonstrated SSIM histogram peak >0.9 between sCT and reference CT. Plan delivery on the Elekta VERSA HD linac proceeded without issues.

Conclusion: AI-generated sCT images provided high geometric accuracy, structural integrity, and dosimetric accuracy for brain and pelvis MRI-only radiotherapy. The successful execution of clinical workflows using sCT-based reference images demonstrates its feasibility for clinical implementation. These findings support the adoption of sCT in MRI-only workflows, enabling radiation therapy without conventional CT while maintaining precise and efficient treatment delivery.