Dixon-based thorax synthetic CT generation using Generative Adversarial Network

Purpose: Generation of synthetic Computed Tomography (sCT) images from Magnetic Resonance (MR) is an imperative, yet not fully resolved task for attenuation correction in Positron Emission Tomography (PET)/MR and treatment planning in MR-only radiation therapy. Herein, we propose a Generative Adversarial Networks (GAN) model to generate quantitatively accurate sCT from Dixon thorax MR data. Materials/methods: Paired image sets of Dixon MR and corresponding CT volumes from fourteen subjects were used. In-phase (IP) MR images were registered to the CT images using REGGUI. The resulting deformation field was applied to the three other [opposed-phase (OP), fat, water] MR images. Image processing was implemented in MATLAB 2016b (MathWorks) using COMKAT Image Tool. Manual contouring was performed on the CT images using MIM (version 6.6.10, MIM Software Inc., Cleveland, OH) software, to delineate the left lung, right lung, vertebral body, and spinal cord. The proposed RU-cGAN model incorporates ResNet and U-Net in the generator of a conditional GAN. Results from Vgg16, Vgg19, and ResNet were used as references for comparison. The four networks were trained using two strategies. In the first, the four Dixon MR images were used as input features. In the second strategy, only fat and water images were used. Results: Regardless of whether the input consists of two or four images types, sCT generated using RU- cGAN had the best agreement with the measured CT for all the metrics, having Mean Absolute Prediction Error (MAPE) ​< ​80 Hounsfield Units and generated in 2.1–2.85 ​s. At the level of the lungs, MAPE and Root Mean Squared Error using four inputs yielded statistically better results when compared to 2 inputs. Conclusion: RU-cGAN provides a rapid and accurate method for thorax sCT generation while obviating the bone-specific UTE pulse sequence and only requiring limited training data.