Magnetic resonance parameter mapping using model-guided self-supervised deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Quantitative MRI parameter mapping is crucial for diagnosing various medical conditions.
• Traditional methods often require fully sampled data, limiting acquisition speed and increasing scan times.
• Accelerated MRI techniques are essential for improving patient comfort and throughput.

Purpose of the Study:

• To develop a novel model-guided, self-supervised deep learning framework named RELAX (reference-free latent map extraction).
• To enable rapid quantitative MR parameter mapping using undersampled k-space data.
• To eliminate the need for fully sampled reference datasets in MRI reconstruction.

Main Methods:

• RELAX incorporates inherent MR imaging and quantitative fitting models for network training.
• Physical model constraints are enforced, allowing direct reconstruction of MR parameter maps from undersampled data.
• Optional inclusion of sparsity constraints, such as total variation, enhances reconstruction quality.

Main Results:

• RELAX successfully generated accurate T1 and T2 maps from simulated and in vivo data, even with noise and undersampling artifacts.
• Spatial total variation constraints improved image quality in reconstructions.
• In vivo performance showed superior reconstruction quality compared to conventional iterative methods and comparable results to supervised deep learning.

Conclusions:

• The study demonstrates the feasibility of rapid quantitative MR parameter mapping using self-supervised deep learning.
• The RELAX framework shows promise for accelerating quantitative MRI applications.
• Future extensions of RELAX can incorporate diverse quantitative imaging models for broader applicability.