Room: Stars at Night Ballroom 2-3
Purpose: The aim of this study was to generate synthetic CT (sCT) image from multi-sequence MR images using an adversarial network for a MR-only brain radiotherapy. We also generated the sCT from only a single sequence and evaluated the dosimetric impacts of the method used for generation of sCT images.
Methods: We evaluated 580 CTâ€“MR image pairs of 15 glioblastoma patients. The sCT was generated using a conditional generative adversarial network (cGAN). For MR images, the T1-weighted (T1w), T2-weighted (T2w), and fluid-attenuated inversion recovery (FLAIR) sequences were used. The sCT images were generated from all three sequences (sCT(multi)) and from a single T1w images (sCT(single)). Image quality was evaluated the voxel-wise mean absolute error (MAE) between sCT and the ground truth CT with respect to the whole body, soft tissue, and bone regions. We also evaluated the volumetric modulated arc therapy treatment plans generated on the sCT images. We analyzed the doseâ€“volume histogram (DVH) and dose distributions by gamma evaluation with a criterion of 1%/1mm.
Results: The mean MAEs of the whole body were 108.1 Â± 24.0 HU and 120.1 Â± 20.4 HU (P<0.01) for sCT(multi) and sCT(single), respectively. For both the sCT(multi) and sCT(single), the D2%, D50%, and D98% relative to the prescribed dose were <1.0% from those of the reference plans. For gamma evaluation, the pass rate of sCT(multi) was 95.3%, which was significantly higher than that of sCT(single) of 94.2% (P=0.022).
Conclusion: The CT number of the sCT(multi) showed better agreement with the original CT than that of sCT(single). Although both methods showed similar dose distributions, the dose calculated on the sCT(multi) showed better similarity to that calculated on the ground truth CT images.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by JSPS KAKENHI Grant number 17K15802.