Purpose: Automatic and accurate generation of synthetic CT (sCT) from abdominal MRI is challenging primarily due to the difficulty to accurately identify randomly occurring air volumes in the abdomen. This work aims to develop an automated deep learning (DL) method to first delineate air volumes on abdominal MRI and then create sCT using air volume overrides.
Methods: The two-step method to create sCT includes: (1) auto-segmenting air volume using a 3D deep convolutional neural network trained with 21 daily abdomen MRI sets acquired on a 1.5T MR-Linac along with manual air volume contours, and (2) creating sCT using a generative adversarial network trained with the same daily MRI sets and overriding and correcting air volumes on the sCT with the air volumes from the first step and an additional CT abdomen segmentation model. The daily MRI sets were acquired with “AirScan”, a 3D Cartesian FLASH sequence with 1 msec echo time, to accentuate the difference between air and non-air tissues. The obtained DL air auto-segmentation model was tested on 7 abdominal AirScan images and compared to manual delineations using Dice similarity coefficient (DSC). Dose calculations (Monaco, Elekta) with Monte Carlo statistical uncertainty of 0.5% and 1.5T magnetic field, using DL vs. manual air volumes were compared using an air density of 0.01 g/cc with gamma analysis and dose-volume histogram (DVH) criteria of important critical structures and targets.
Results: The execution time for the model was less than 2 seconds. The DSC between DL and manual air contours was 98% ±1%. The gamma passing rate (1% - 1mm criteria) was 99%± 1%, and differences in DVH criteria were <1%, and generally smaller than differences due to statistical uncertainty
Conclusion: The new DL air auto-segmentation model provided accurate air volumes for DL sCT generation in abdomen, facilitating MR-guided radiotherapy for abdominal malignancies.
Funding Support, Disclosures, and Conflict of Interest: The research was partially supported by the Medical College of Wisconsin (MCW) Cancer Center and Froedtert Hospital Foundation, the MCW Meinerz and Fotsch Foundations, and the National Cancer Institute of the National Institutes of Health under award number R01CA247960.
Not Applicable / None Entered.