Generating synthetic CTs from magnetic resonance images using generative adversarial networks

PURPOSE: While MR-only treatment planning using synthetic CTs (synCTs) offers potential for streamlining clinical workflow, a need exists for an efficient and automated synCT generation in the brain to facilitate near real-time MR-only planning. This work describes a novel method for generating brain synCTs based on generative adversarial networks (GANs), a deep learning model that trains two competing networks simultaneously, and compares it to a deep convolutional neural network (CNN). METHODS: Post-Gadolinium T1-Weighted and CT-SIM images from fifteen brain cancer patients were retrospectively analyzed. The GAN model was developed to generate synCTs using T1-weighted MRI images as the input using a residual network (ResNet) as the generator. The discriminator is a CNN with five convolutional layers that classified the input image as real or synthetic. Fivefold cross-validation was performed to validate our model. GAN performance was compared to CNN based on mean absolute error (MAE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR) metrics between the synCT and CT images. RESULTS: GAN training took ~11 h with a new case testing time of 5.7 ± 0.6 s. For GAN, MAEs between synCT and CT-SIM were 89.3 ± 10.3 Hounsfield units (HU) and 41.9 ± 8.6 HU across the entire FOV and tissues, respectively. However, MAE in the bone and air was, on average, ~240-255 HU. By comparison, the CNN model had an average full FOV MAE of 102.4 ± 11.1 HU. For GAN, the mean PSNR was 26.6 ± 1.2 and SSIM was 0.83 ± 0.03. GAN synCTs preserved details better than CNN, and regions of abnormal anatomy were well represented on GAN synCTs. CONCLUSIONS: We developed and validated a GAN model using a single T1-weighted MR image as the input that generates robust, high quality synCTs in seconds. Our method offers strong potential for supporting near real-time MR-only treatment planning in the brain