Development and clinical implementation of an MRI-only planning workflow featuring deep learning-based synthetic CT...

In external beam radiation therapy (EBRT) for prostate cancer, both MRI and CT are typically used-CT provides electron density for dose calculation and visualization of bony anatomy and fiducial markers, while MRI offers superior soft tissue contrast. With recent advances in deep learning, synthetic CT (sCT) images can now be generated from MRI, potentially eliminating the need for separate CT scans.

To clinically implement an MRI-only planning (MROP) workflow for both X-ray and MRI-guided systems.

MROP implementation involved optimizing MRI simulation protocols, developing a deep learning-based sCT generation method, and automating fiducial marker detection. A multichannel CycleGAN model was trained using T1-weighted Dixon images to generate sCT. For SBRT cases, the immobilization device was auto-inserted into sCT images. Fiducial markers were detected using a quantitative susceptibility mapping (QSM) method. Validation included retrospective comparison of Hounsfield units and dosimetric plans to those based on planning CT for 11 patients, followed by prospective evaluation of ten patients.

The MRI protocol included T2-weighted 3D imaging for contouring, T1-weighted Dixon for sCT, and T2*-weighted GRE for fiducial detection. The MROP workflow was prospectively applied in 10 patients (three SBRT, seven conventional). Dosimetric comparisons showed <1% difference between MROP and CT-based plans across all PTVs. The QSM-based fiducial detection achieved a 95.5% success rate (63/66 markers) and reduced detection time from ∼10 to 2 min per patient. All MROP components were integrated into the treatment planning system via scripting. Implementation reduced simulation scheduling time by 7.1 days.

We successfully implemented an MROP workflow for prostate cancer EBRT, eliminating the need for CT, reducing radiation exposure, and improving clinical efficiency. Over 500 patients have been treated to date using this approach on both C-arm and MR-guided systems.