Room: Karl Dean Ballroom B1
Purpose: Clinical implementation of MRI guided radiotherapy requires a method to derive synthetic CT (S-CT) for dose calculation. This study aims to investigate the feasibility of building a deep convolutional neural network for MRI based S-CT generation and evaluate the dosimetry accuracy on prostate IMRT planning.
Methods: A Paired CT and T2-weighted MR images were acquired from each of 36 prostate cancer patients. Ten CT-MR pairs were used as tested set and the other 26 pairs as training set. The training set were augmented by applying 3-dimensional artificial deformations and then feed to a 2-dimensional U-net which contains 23 convolutional layers and 25.29 million trainable parameters. The Hounsfield unit (HU) accuracy of the U-net generated S-CT images were evaluated by the mean absolute error (MAE) and compared to that of those S-CT images generated by a multi-atlas deformable image registration based method. Prostate IMRT plans were applied on the true CT images. The true CT images were replaced by the S-CT images and the dose matrices were recalculated and compared to the one that obtained from the true CT using 3-dimensional global Î³-index.
Results: The U-net was trained in 38 hours on a Quadro M6000 GPU with 12GB memory. The computation time for generating an S-CT volume image was 6.68~9.09 seconds. The (mean Â± standard deviation) MAE within body was (31.52Â±6.16)HU and (40.16Â±4.75)HU respect to the U-net and the multi-atlas method (p<0.002). Within the region of interests, the U-net generated S-CT images achieved the (1mm&1%) Î³ pass rate over 99.91%, the maximum point dose difference less than 1.1% respect to the prescription and the discrepancy of dose-volume parameters less than 0.52%.
Conclusion: The U-net can generate accurate S-CT volume images within seconds with the maximum point dose error less than 1.1% within the region of interests for prostate IMRT plan.