RAIDER: Rapid, anatomy-independent deep learning-based chemical shift-encoded MRI

Purpose: Despite recent advances, chemical shift-encoded MRI (CSE-MRI) remains a challenging problem and many algorithms are computationally expensive, leading to interest in deep learning-based methods. However, initial attempts have used convolutional neural networks (CNNs), which are limited by data requirements, poor generalisability across different anatomies (anatomy-dependence) and training time. To address these limitations, we propose RAIDER, a method for rapid, anatomy-independent deep learning-based CSE-MRI. Theory and Methods: RAIDER uses two multilayer perceptrons (MLPs), each trained separately with simulated single-voxel data, thus avoiding the degeneracy encountered during training when using only one network. During inference, the solution from one of the two networks is chosen based on likelihood. Performance and speed are investigated in a series of simulation experiments, in phantoms and in vivo. Results: RAIDER is approximately 700 times faster than conventional fitting, taking 14us per voxel rather than 10ms, and offers performance similar to that of conventional fitting. It produces accurate fat fraction measurements in phantoms and in vivo images with different anatomies, despite having been trained only on simulation data. Conclusion: RAIDER resolves the problems posed by degeneracy, avoids the training data requirements of CNN-based methods and markedly reduces computational cost compared to conventional fitting.