Purpose: MR scans in MR-guided radiotherapy can be partially truncated due to limited field of view, which could significantly affect dose calculation accuracy for MR-based online adaptive planning. We developed a cycle consistent generative adversarial networks (cycleGAN) based approach to create synthetic CT (sCT) images compensating the missing anatomy from the truncated MR scans.
Methods: CT scans and T1-weighted MR scans with complete anatomy of 24 head and neck patients were used for this study. The MR scans were manually cropped 10-15 mm off at the posterior region of head to simulate truncated MR scans. Six patients were randomly chosen for testing and the rest were used for training. We proposed a modified cycleGAN using a residual-Unet as the generator to create sCT from MR. Truncated/complete MR scan pairs and CT scans were used to train the cycleGAN. The sCT pairs were generated for both truncated and complete MR scans. The mean absolute difference (MAD) of Hounsfield units (HU) was used to quantify the sCT performance. To evaluate the sCT in the truncated region, body contours were generated for each sCT pair. In the truncated region, the structural similarity index (SSIM), mean surface distance (MSD), and Dice similarity of body contours were calculated between sCT pairs to quantify the sCT performance in compensating the truncated anatomy.
Results: The mean MAD were 83.2±9.3 HU and 81.6±9.2 HU for the sCT generated from truncated and complete MRI, respectively. In the truncated region, the mean SSIM was 0.86±0.02, the MSD was 1.91±0.52mm, and the Dice of body contour was 0.92±0.02, indicating a good performance in compensating the truncated anatomy.
Conclusion: We developed a novel cycleGAN model that can create synthetic CT with complete anatomy from truncated MR scans, which could potentially benefit online MR-guided adaptive treatment planning.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by MD Anderson Cancer Center Startup fund