Dixon MRI is widely used for body composition studies. Current processing methods associated with large whole-body volumes are time intensive and prone to artifacts during fat-water separation performed on the scanner, making the data difficult to analyse. The most common artifact are fat-water swaps, where the labels are inverted at the voxel level. It is common for researchers to discard swapped data (generally around 10%), which can be wasteful and lead to unintended biases. The UK Biobank is acquiring Dixon MRI for over 100,000 participants, and thousands of swaps will occur. If those go undetected, errors will propagate into processes such as abdominal organ segmentation and dilute the results in population-based analyses. There is a clear need for a fast and robust method to accurately separate fat and water channels. In this work we propose such a method based on style transfer using a conditional generative adversarial network. We also introduce a new Dixon loss function for the generator model. Using data from the UK Biobank Dixon MRI, our model is able to predict highly accurate fat and water channels that are free from artifacts. We show that the model separates fat and water channels using either single input (in-phase) or dual input (in-phase and opposed-phase), with the latter producing improved results. Our proposed method enables faster and more accurate downstream analysis of body composition from Dixon MRI in population studies by eliminating the need for visual inspection or discarding data due to fat-water swaps.
Download