CYCLE-CONSISTENT SELF-SUPERVISED LEARNING FOR IMPROVED HIGHLY-ACCELERATED MRI RECONSTRUCTION.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Physics

Background:

• Physics-driven deep learning (PD-DL) accelerates MRI acquisition.
• Unsupervised learning, including self-supervised learning, is emerging in PD-DL.
• Current methods degrade performance at very high acceleration rates.

Purpose of the Study:

• To improve self-supervised learning for highly accelerated MRI using cyclic-consistency (CC).
• To reduce aliasing artifacts in accelerated MRI.

Main Methods:

• Proposed a cyclic-consistency approach for self-supervised learning in PD-DL.
• Generated simulated measurements by undersampling network output.
• Used a masking-based self-supervised loss in conjunction with CC.

Main Results:

• Substantially reduced aliasing artifacts at high acceleration rates.
• Demonstrated effectiveness on rate 6 and 8 fastMRI knee imaging.
• Validated performance on 20-fold HCP-style fMRI data.

Conclusions:

• Cyclic-consistency improves self-supervised learning for highly accelerated MRI.
• The proposed method enhances image quality and reduces artifacts.
• This technique is effective for challenging acceleration scenarios.