Deep learning for fast low-field MRI acquisitions.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Biophysics

Background:

• Low-field MRI (LF MRI) offers cost and accessibility benefits but suffers from low signal-to-noise ratio, leading to long acquisition times.
• Deep learning (DL) shows promise for accelerating MRI acquisition through undersampling, but requires large datasets typically unavailable in LF regimes.

Purpose of the Study:

• To demonstrate the efficacy of a Residual U-net model combined with data augmentation for reconstructing undersampled LF MRI scans.
• To evaluate the model's performance using a limited training dataset (n=10) at 0.1 Tesla (T).

Main Methods:

• Utilized a Residual U-net architecture integrated with data augmentation techniques.
• Reconstructed magnitude and phase information from undersampled LF MRI scans.
• Evaluated model performance retrospectively across various acceleration rates and sampling patterns, followed by prospective validation on fivefold undersampled data.

Main Results:

• The DL approach successfully reconstructed magnitude and phase images from undersampled LF MRI data.
• Performance varied with sampling schemes, but the model preserved global structure and image sharpness.
• The method demonstrated capability even with a small training dataset (n=10).

Conclusions:

• Deep learning, particularly Residual U-net with data augmentation, can effectively reconstruct undersampled LF MRI data.
• This approach addresses the challenge of limited training data in LF MRI research.
• The findings suggest a viable pathway for improving LF MRI clinical relevance and implementation.