Improving generalization in MR-to-CT synthesis in radiotherapy by using an augmented cycle generative adversarial network with unpaired data.

Purpose MR-to-CT synthesis is one of the first steps in the establishment of an MRI-only workflow in radiotherapy. Current MR-to-CT synthesis methods in deep learning use unpaired MR and CT training images with a Cycle Generative Adversarial Network (CycleGAN) to minimize the effect of misalignment between paired images. However, this approach critically assumes that the underlying inter-domain mapping is approximately deterministic and one-to-one. In the current study, we use an Augmented CycleGAN (AugCGAN) model to create a robust model that can be applied to different scanners and sequences using unpaired data. Materials and methods This study included T2-weighted MR and CT pelvic images of 38 patients in treatment position from 5 different centers. The AugCGAN was trained on 2D transverse slices of 19 patients from 3 different sites. The network was then used to generate synthetic CT (sCT) images of 19 patients from the two other sites. Mean absolute errors (MAEs) for each patient were evaluated between real and synthetic CT images. Original treatment plans of 9 patients were re-calculated using sCT images to assess the dose distribution in terms of voxel-wise dose difference, gamma, and dose volume histogram analysis. Results The mean MAEs were 59.8 Hounsfield units (HU) and 65.8 HU for the first and second test sites, respectively. The maximum dose difference to the target was 1.2% with a gamma pass rate using the 3%, 3 mm criteria above 99%. The average time required to generate a complete sCT image for a patient on our GPU was 8.5 s. Conclusion This study suggests that our unpaired approach achieves good performance in generalization with respect to sCT image generation.