CT Synthesis from MRI using 3D Swin UNETR and Distillation for Upper Abdominal Radiotherapy Treatment Planning

Deep learning-based synthetic CT from MRI could greatly simplify the workflow in MR-based radiotherapy planning and provide more accurate GTV delineation thanks to the better tumor contrast in MRI. Model development for the upper abdomen remains challenging due to the large organ deformation between the CT and MRI. In this work, we proposed a novel distillation method for deep learning-based CT synthesis from MRI. A 3D Swin UNETR was trained and tuned on 192 paired CT and Dixon MR images, followed by postprocessing steps to correct artifacts such as diminishing skins to generate initial synthetic CTs. A distilled Swin UNETR was then trained to map the MRI to the initial synthetic CTs to integrate the postprocessing steps into a single deep neural network. The model was tested on 21 patients with both planning MRI and CT acquired prospectively. The mean absolute error (MAE) between synthetic and planning CT was 62.39 HU. The clinical treatment plans were recomputed on the synthetic CTs. Compared to the planning CT, the synthetic CT achieved a mean PTV dose error of 2.34% and a mean 3%/3mm gamma passing rate of 96.23%.