Improving Fast MRI Reconstructions with Pretext Learning in Low-Data Regime.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Supervised deep learning accelerates magnetic resonance (MR) imaging but requires extensive labeled data.
• Acquiring fully-sampled raw MR data for training is costly and time-consuming.
• Existing methods often use retrospectively under-sampled data, simulating accelerated scans.

Purpose of the Study:

• To introduce a self-supervised learning pretext method to improve MR image reconstruction.
• To leverage commonly available under-sampled MR data for feature learning.
• To enhance deep learning models in low-data regimes for MR imaging.

Main Methods:

• Developed a self-supervised learning pretext task to boost feature learning.
• Utilized under-sampled MR data as the primary training resource.
• Implemented and evaluated the method across various deep-learning-based reconstruction models.

Main Results:

• Self-supervision demonstrated stable training in low-data scenarios.
• Improved MR image reconstruction quality was observed.
• The pretext method effectively boosted feature learning from under-sampled data.

Conclusions:

• Self-supervised learning offers a viable solution to the data demands of supervised deep learning in MR imaging.
• This approach enhances the efficiency and stability of MR image reconstruction models.
• It enables high-fidelity reconstructions even when fully-sampled data is scarce.