A model-based MR parameter mapping network robust to substantial variations in acquisition settings.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Quantitative MRI

Background:

• Deep learning excels at magnetic resonance (MR) parameter mapping (MPM) but typically requires specific acquisition settings.
• Real-world MR scans vary significantly across centers, scanners, and studies.
• A need exists for deep learning MPM methods that are robust to these acquisition variations.

Purpose of the Study:

• To develop a robust deep learning model for quantitative MR parameter mapping adaptable to varying acquisition settings.
• To address the limitations of current MPM methods that are sensitive to changes in scan protocols.

Main Methods:

• Developed MMPM-Net, a model-based deep network integrating a deep learning denoiser into the nonlinear inversion problem of MPM.
• Utilized the alternating direction method of multipliers (ADMM) to solve the optimization problem and unroll it into the network architecture.
• Incorporated a data fidelity component to handle variations in acquisition parameters.

Main Results:

• MMPM-Net demonstrated robust performance on R2 and R1 mapping datasets with significant variations in acquisition settings.
• Qualitative and quantitative experiments showed MMPM-Net outperformed state-of-the-art MR parameter mapping methods.
• The method effectively handles variations in acquisition parameters, a key challenge in practical MR imaging.

Conclusions:

• MMPM-Net offers a robust and versatile solution for quantitative MR parameter mapping across diverse acquisition settings.
• The proposed approach advances the applicability of deep learning in clinical MR imaging by accommodating protocol variability.
• This work paves the way for more reliable and widespread use of deep learning-based MPM in practice.