Improved Robustness in Water-Fat Separation in MRI using Conditional Adversarial Networks

Water-fat separation is a post-processing method to obtain water/fat only images and parametric maps from multi-echo magnetic resonance (MR) images. Due to multi-parametric analytic models and optimization algorithm, the water-fat separation problem is complicated and time-consuming to solve. Traditional model-based techniques require a known field map to make the problem becomes “almost linear”, which results in the dependence on the accuracy of field map estimation and the decrease of computing efficiency. In this study, we proposed a deep learning based method to solve the inverse problem and simultaneously obtain the water/fat images, field map and R2* map without iteration process and field map estimation in advance. Conditional GAN was utilized in this work to preserve the structural details and ground truth was obtained using a graph cut method. The results showed that our method had a more robust performance and higher structural similarity in water-fat separation compared to U-Net based method. The proposed deep learning method is field map free and effective to separate fat/water.