RAIDER-SSL: Self-supervised machine learning for rapid, anatomy-independent PDFF and R2* estimation using magnitude-only signals

Motivation: RAIDER[1] enables voxel-wise, rapid, anatomy-independent PDFF and R2* estimation using supervised learning. However, supervised learning requires target data distribution known a priori; its performance may suffer if out-of distribution data is encountered. Self-supervised learning (SSL) has been proposed to address this limitation. Goal(s): To develop a self-supervised alternative to RAIDER. Approach: The dual network approach of RAIDER is adapted to SSL. This is compared against the standard SSL which uses a single network. Results: RAIDER-SSL outperforms the standard SSL implementation, both in simulation and in vivo. Impact: RAIDER-SSL enables PDFF and R2* estimation from voxel-wise, magnitude-only CSE-MRI data while avoiding the potential performance loss associatedwith distributional shift that may appear with supervised learning methods.