Multi-Site Validation of a Universal Deep Learning–Based Synthetic CT Model for MR-Only Radiotherapy
Abstract
Purpose
MRI–only radiotherapy planning requires accurate synthetic CT (sCT) generation to enable dose calculation and patient positioning without a planning CT in Linac-based treatment delivery settings. While prior studies have demonstrated promising results for individual anatomical sites, robust multi-site solutions remain limited. This study develops and comprehensively evaluates a universal deep learning based sCT generation model for MR-only radiotherapy across multiple anatomical sites, including head and neck, thorax, abdomen, pelvis, and spine.
Methods
MRI, planning CT, and treatment-day CBCT data from 164 patients encompassing head and neck, thoracic, abdominal, pelvic, and spinal sites were retrospectively analyzed. A universal deep learning framework was trained using mDixon in-phase and water MRI sequence as inputs to generate sCT images applicable to all sites. Performance was assessed through three complementary evaluations: voxel-wise HU agreement relative to CT, dose recalculation accuracy of clinical treatment plans on sCT, and consistency of rigid CBCT-based patient alignment when using sCT versus CT as the reference image.
Results
The universal model produced anatomically consistent sCTs across all regions, with HU discrepancies lowest in soft tissue and higher in bony structures. Despite these differences, dose recalculations on sCT showed strong agreement with CT-based plans, with target dose deviations generally below 1 Gy and gamma pass rates above 95% (2 mm/2%). Patient alignment based on sCT demonstrated close agreement with CT-based alignment for head and neck and spine cases, while abdominal and pelvic sites exhibited increased variability driven by soft-tissue motion and MRI-related artifacts rather than sCT generation errors.
Conclusion
A single deep learning–based sCT model can support MR-only radiotherapy across multiple anatomical sites with clinically acceptable accuracy in imaging, dosimetry, and positioning. These findings suggest that a universal sCT framework is a viable path toward scalable MR-only workflows, provided that site-specific quality assurance is applied in anatomically complex regions.