Recurrent inference machines as inverse problem solvers for MR relaxometry.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Quantitative MRI (QMRI)

Background:

• Quantitative MRI (QMRI) is crucial for accurate tissue characterization.
• Conventional methods like Maximum Likelihood Estimator (MLE) can be computationally intensive.
• Recurrent Inference Machines (RIMs) have shown promise in solving inverse problems.

Purpose of the Study:

• To propose and evaluate Recurrent Inference Machines (RIMs) for T1 and T2 mapping in QMRI.
• To demonstrate RIMs' capability in optimizing non-linear problems for accurate relaxometry.
• To compare RIM performance against MLE and Residual Neural Network (ResNet) methods.

Main Methods:

• Development of a RIM framework for iterative inference based on MRI signal models.
• Evaluation using simulated data, phantom studies, and in-vivo human scans.
• Comparison of accuracy, precision, and computational speed against MLE and ResNet.

Main Results:

• RIMs demonstrated improved accuracy and precision in T1 and T2 mapping compared to MLE and ResNet.
• RIM inference was found to be approximately 150 times faster than MLE.
• The RIM framework showed robustness to minor variations in scanning parameters.

Conclusions:

• Recurrent Inference Machines are a promising, flexible, and efficient tool for QMRI relaxometry.
• RIMs offer a compelling alternative to existing QMRI methods, combining data-driven and model-based advantages.
• The availability of open-source training tools further enhances the utility of RIMs in QMRI.