SuperMAP: Deep ultrafast MR relaxometry with joint spatiotemporal undersampling.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Quantitative MRI

Background:

• Quantitative magnetic resonance imaging (MRI) provides valuable tissue-specific information but is often limited by long acquisition times.
• Developing faster methods for MR parameter mapping is crucial for clinical applications and reducing patient discomfort.

Purpose of the Study:

• To develop an ultrafast and robust deep learning-based network for magnetic resonance (MR) parameter mapping.
• To enable simultaneous reconstruction of T1rho and T2 relaxation maps within a single scan.

Main Methods:

• A deep learning framework, SuperMAP, was designed to directly reconstruct quantitative maps from undersampled parameter-weighted images, bypassing traditional fitting procedures.
• The network was trained and tested using retrospectively undersampled data and validated on prospectively acquired data.
• Performance was evaluated using parameter qualification errors and similarity metrics against gold-standard methods.

Main Results:

• SuperMAP achieved accurate T1rho and T2 mapping at high acceleration factors (R=24 and R=32).
• The method demonstrated low normalized mean square error (2.7%-2.8%) and high structural similarity (96%-97%) compared to gold standards.
• The network trained on retrospective data performed well on prospective data, outperforming conventional methods.

Conclusions:

• Deep learning enables the generation of superfast MR parameter maps from highly undersampled data.
• SuperMAP offers a feasible approach for simultaneous T1rho and T2 relaxation mapping in significantly reduced scan times.