Spatially regularized parametric map reconstruction for fast magnetic resonance fingerprinting

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Quantitative MRI

Background:

• Magnetic Resonance Fingerprinting (MRF) enables rapid acquisition of multiple quantitative MRI parameters.
• Current MRF reconstruction relies on dictionary matching, which is computationally intensive and limits clinical scalability.
• There is a need for faster, more accurate, and scalable MRF reconstruction methods.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN)-based reconstruction for MRF.
• To enable accurate and fast simultaneous mapping of T1 relaxation time of water (T1_H2O) and fat fraction (FF).
• To assess the CNN's performance on a diverse patient cohort and its generalization capabilities.

Main Methods:

• A CNN-based reconstruction framework was developed for MRF data.
• The method was evaluated using the MRF T1-FF sequence for T1_H2O and FF mapping.
• The CNN incorporated spatial regularization and was trained and tested on a heterogeneous dataset of 164 patients with neuromuscular diseases.

Main Results:

• The CNN reconstruction achieved high accuracy, outperforming state-of-the-art deep learning methods.
• Normalized root mean squared errors were 0.048 ± 0.011 for T1_H2O and 0.027 ± 0.004 for FF maps compared to dictionary matching.
• The method demonstrated robustness in handling heterogeneous morphometric variations and generalized to unseen anatomical regions.

Conclusions:

• The proposed CNN-based MRF reconstruction offers accurate and fast parametric mapping.
• This method addresses the computational and scalability challenges of dictionary matching.
• It holds significant potential for enabling efficient multiparametric MRI in clinical practice.