Flexible and cost-effective deep learning for accelerated multi-parametric relaxometry using phase-cycled bSSFP.

Area of Science:

• Medical Imaging
• Machine Learning
• Neuroscience

Background:

• Quantitative magnetic resonance imaging (qMRI) adoption is hindered by slow acquisition and complex analysis.
• Accurate parameter mapping is crucial for clinical applications of qMRI.
• Developing efficient and flexible qMRI frameworks is essential for widespread clinical use.

Purpose of the Study:

• To compare deep neural network (DNN) and iterative fitting frameworks for multi-parametric (MP) relaxometry using phase-cycled balanced steady-state free precession (pc-bSSFP) imaging.
• To evaluate the performance of supervised DNNs (SVNN), physics-informed DNNs (PINN), and MIRACLE in silico and in vivo.
• To assess the potential of DNNs for accelerating data acquisition and improving the robustness of qMRI.

Main Methods:

• Comparison of SVNN, PINN, and MIRACLE frameworks for MP relaxometry.
• In silico and in vivo evaluation using brain tissue from healthy subjects.
• Monte Carlo sampling for noise simulation and DNN training on diverse parameter distributions and signal-to-noise ratios.
• Utilizing complex-valued MR data for DNN-based acceleration.

Main Results:

• PINNs demonstrated superior consistency and robustness to training data variations compared to SVNNs.
• DNNs accelerated data acquisition by a factor of 3 by leveraging complex-valued MR data.
• Whole-brain relaxometry using DNNs was effective, adaptive, and showed potential for low-cost retraining.
• In silico DNN pipelines enabled rapid data generation and training without extensive dictionaries or long inference times.

Conclusions:

• DNN-based MP-qMRI pipelines offer a flexible and rapid approach for quantitative imaging.
• Physics-informed DNNs enhance the reliability and adaptability of qMRI parameter mapping.
• This work highlights the advantages of lightweight machine learning for accelerating clinical adoption of qMRI.