A spatiotemporal dependency-aware lightweight CNN-ViT network for 3D MRF with a balanced acceleration strategy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Rapid Magnetic Resonance Imaging (MRI) acquisition is crucial for clinical efficiency and diagnostic consistency.
• 3D Magnetic Resonance Fingerprinting (MRF) offers fast, multi-parametric quantitative imaging but faces limitations in accuracy and scan time.
• Existing deep learning methods for accelerating MRF have limitations in acceleration strategy and modeling complex spatiotemporal data.

Purpose of the Study:

• To develop an accelerated 3D MRF technique that preserves quantitative accuracy for clinical adoption.
• To address limitations in current deep learning approaches for MRF acceleration.
• To enable faster, whole-brain MRF imaging with improved accuracy and image quality.

Main Methods:

• Proposed a lightweight spatiotemporal attention enhanced network (LiST-UNet) integrating CNNs and Vision Transformer components.
• Incorporated a precursor-successor network to model interrelationships between tissue parameters for enhanced T2 quantification.
• Implemented a balanced k-space and temporal-frame acceleration strategy to minimize reconstruction errors.

Main Results:

• Achieved whole-brain 3D MRF imaging in approximately 1.25 minutes, an eightfold acceleration.
• Demonstrated superior quantification accuracy and image quality compared to previous deep learning methods.
• The balanced acceleration strategy significantly reduced errors compared to single-dimension undersampling.

Conclusions:

• The LiST-UNet enables significantly faster 3D MRF acquisition without compromising quantitative accuracy.
• This method overcomes limitations of existing deep learning approaches for MRF acceleration.
• The combined architectural improvements and acceleration strategy support the clinical translation of 3D MRF.