RAIDER: Rapid, anatomy-independent, deep learning-based PDFF and R2* estimation using magnitude-only signals, dual neural networks and training data distribution design

There has been recent interest in the use of magnitude-based fitting methods for estimating proton density fat fraction (PDFF) and R2∗ from chemical shift-encoded MRI (CSE-MRI) data, since these methods can still be used when complex based methods fail or when phase data are inaccessible or unreliable (such as in multicentre trials, low resource settings, preclinical imaging and national cohort datasets), and may also be used as a final processing step with complex-based methods. However, conventional fitting techniques are computationally expensive. Deep learning (DL)-based methods promise to accelerate parameter estimation, but there are no existing deep learning methods for voxelwise CSE-MRI tissue parameter estimation. Here, we show that a naive voxelwise multilayer perceptron (MLP) implementation suffers from poor performance because multiple tissue parameter values can produce similar signals (degeneracy), potentially accounting for this gap in the literature. To address this problem and realise the potential acceleration offered by MLP-based tissue parameter estimation, we propose RAIDER, a voxelwise method for rapid, anatomy-independent deep learning-based PDFF and R2∗ estimation using multi-echo magnitude-data. RAIDER utilises two neural networks, each with a separately-designed training data distribution, to deal with degeneracy and thus realise the benefits of an MLP-based approach, with a 400-fold to 2800-fold acceleration (meaning that processing takes just seconds, rather than minutes or hours). Our code is available at https://github.com/TJPBray/DixonDL