The establishment of an MRI-only workflow in radiotherapy depends on the ability to generate an accurate synthetic CT (sCT) for dose calculation. Previously proposed methods have used a Generative Adversarial Network (GAN) for fast sCT generation in order to simplify the clinical workflow and reduces uncertainties. In the current paper we use a conditional Generative Adversarial Network (cGAN) framework called pix2pixHD to create a robust model prone to multicenter data. This study included T2-weighted MR and CT images of 19 patients in treatment position from 3 different sites. The cGAN was trained on 2D transverse slices of 11 patients from 2 different sites. Once trained, the network was used to generate sCT images of 8 patients coming from a third site. The Mean Absolute Errors (MAE) for each patient were evaluated between real and synthetic CTs. A radiotherapy plan was optimized on the sCT series and re-calculated on CTs to assess the dose distribution in terms of voxel-wise dose difference and Dose Volume Histograms (DVH) analysis. It takes on average of to generate a complete sCT (88 slices) for a patient on our GPU. The average MAE in HU between the sCT and actual patient CT (within the body contour) is 48.5 ± 6 HU with our method. The maximum dose difference to the target is 1.3%. This study demonstrates that an sCT can be generated in a multicentric context, with fewer pre-processing steps while being fast and accurate.

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 interdomain 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 five different centers. The AugCGAN was trained on 2D transverse slices of 19 patients from three 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 nine patients were recalculated 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 (italicHU) 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.