(Thursday, 8/2/2018) 1:00 PM - 3:00 PM
Room: Karl Dean Ballroom B1
Purpose: Synthetic CTs (sCTs) are used for dose calculation and patient positioning in MR-only radiotherapy. In this work, we developed a conditional generative adversarial network (cGAN) for fast abdominal sCT generation using a single low-field MR scan, and evaluated sCT dosimetric accuracy for MR-only liver radiotherapy.
Methods: A retrospective study was performed using CTs and 0.35 T MR images from 12 abdominal cancer patients. All CTs were deformably registered to MR images to generate paired MR-CT slices for model training. The proposed cGAN is based on the “Pix2Pix� network with the generator replaced with a 27-layer convolutional neural network. The 12-patient cohort was randomly divided into 4 groups of 3 patients. sCTs of patients in each group were generated by the cGAN trained on the other 3 groups. 8 liver patients were used to test the sCT dosimetric accuracy. Volumetric modulated arc therapy (VMAT) plans were optimized on CTs according to standard clinical guidelines and recalculated on the corresponding sCTs. CT and sCT plans were compared using mean dose difference, dose-volume histogram parameters, and gamma analyses.
Results: On average it takes 10.5 s to generate each sCT. The overall absolute mean dose deviations between CT and sCT dose distributions were 0.25 Gy for planning target volumes (PTVs) and below 0.04 Gy for all evaluated OARs. The mean gamma passing rates were 99.7±0.6% for PTV and 99.9±0.2% inside the body using a 3%/3-mm criterion.
Conclusion: Results show that abdominal sCTs generated by the proposed cGAN achieve high dosimetric accuracy for liver radiotherapy. The high dosimetric accuracy and fast generation speed achieved by the cGAN make it a promising tool for MR-only abdominal radiotherapy and MR-guided online adaptive planning. More abdominal patients will be enrolled to further compare CT and sCT plans using statistical equivalence tests.
Funding Support, Disclosures, and Conflict of Interest: This study was partially funded by the Varian master research agreement.
Keywords
Low-field MRI, Image-guided Therapy, Computer Vision
Taxonomy
IM/TH- MRI in Radiation Therapy: MRI for treatment planning
Contact Email
