RESUMO
PURPOSE: The first aim was to generate and compare synthetic-CT (sCT) images using a conditional generative adversarial network (cGAN) method (Pix2Pix) for MRI-only prostate radiotherapy planning by testing several generators, loss functions, and hyper-parameters. The second aim was to compare the optimized Pix2Pix model with five other architectures (bulk-density, atlas-based, patch-based, U-Net, and GAN). METHODS: For 39 patients treated by VMAT for prostate cancer, T2-weighted MRI images were acquired in addition to CT images for treatment planning. sCT images were generated using the Pix2Pix model. The generator, loss function, and hyper-parameters were tuned to improve sCT image generation (in terms of imaging endpoints). The final evaluation was performed by 3-fold cross-validation. This method was compared to five other methods using the following imaging endpoints: the mean absolute error (MAE) and mean error (ME) between sCT and reference CT images (rCT) of the whole pelvis, bones, prostate, bladder, and rectum. For dose planning analysis, the dose-volume histogram metric differences and 3D gamma analysis (local, 1 %/1 mm) were calculated using the sCT and reference CT images. RESULTS: Compared with the other architectures, Pix2Pix with Perceptual loss function and generator ResNet 9 blocks showed the lowest MAE (29.5, 107.7, 16.0, 13.4, and 49.1 HU for the whole pelvis, bones, prostate, bladder, and rectum, respectively) and the highest gamma passing rates (99.4 %, using the 1 %/1mm and 10 % dose threshold criterion). Concerning the DVH points, the mean errors were -0.2% for the planning target volume V95%, 0.1 % for the rectum V70Gy, and -0.1 % for the bladder V50Gy. CONCLUSION: The sCT images generated from MRI data with the Pix2Pix architecture had the lowest image errors and similar dose uncertainties (in term of gamma pass-rate and dose-volume histogram metric differences) than other deep learning methods.
