SPARSITY-DRIVEN PARALLEL IMAGING CONSISTENCY FOR IMPROVED SELF-SUPERVISED MRI RECONSTRUCTION.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Physics-driven deep learning (PD-DL) models enhance rapid MRI reconstruction.
• Self-supervised learning (SSL) is crucial for training when fully-sampled data is absent.
• High acceleration rates in MRI often lead to artifacts and reduced image fidelity with SSL.

Purpose of the Study:

• To propose a novel training strategy for PD-DL networks to improve MRI reconstruction quality.
• To address the challenge of artifacts and noise amplification in self-supervised MRI at high acceleration rates.
• To develop a method that enhances k-space masking with a novel consistency term for artifact reduction.

Main Methods:

• Developed a novel training strategy for PD-DL networks using carefully designed perturbations.
• Enhanced k-space masking with a new consistency term to predict added perturbations in a sparse domain.
• Validated the approach on fastMRI knee and brain datasets.

Main Results:

• The proposed method effectively reduces aliasing artifacts in accelerated MRI.
• Demonstrated mitigation of noise amplification at high acceleration rates.
• Outperformed existing state-of-the-art self-supervised methods in visual and quantitative assessments.

Conclusions:

• The novel training strategy leads to more reliable and artifact-free MRI reconstructions.
• This approach significantly improves image fidelity for rapid MRI scans using self-supervised learning.
• The method offers a promising solution for high-acceleration MRI where reference data is unavailable.