Ground-truth free multi-mask self-supervised physics-guided deep learning in highly accelerated MRI

Deep learning based MRI reconstruction methods typically require databases of fully-sampled data as reference for training. However, fully-sampled acquisitions may be either challenging or impossible in numerous scenarios. Self-supervised learning enables training neural networks for MRI reconstruction without fully-sampled data by splitting available measurements into two disjoint sets. One of them is used in data consistency units in the network, and the other is used to define the loss. However, the performance of self-supervised learning degrades at high acceleration rates due to scarcity of acquired data. We propose a multi-mask self-supervised learning approach, which retrospectively splits available measurements into multiple 2-tuples of disjoint sets. Results on 3D knee and brain MRI shows that the proposed multi-mask self-supervised learning approach significantly improves upon single mask self-supervised learning at high acceleration rates.