Two-stage deep learning for accelerated 3D time-of-flight MRA without matched training data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Time-of-flight magnetic resonance angiography (TOF-MRA) is crucial for non-contrast vascular imaging.
• Accelerated acquisition is essential for 3D TOF-MRA, leading to undersampled data.
• Supervised deep learning for TOF-MRA reconstruction requires difficult-to-obtain matched reference data.

Purpose of the Study:

• To develop a high-quality reconstruction method for undersampled TOF-MRA using unsupervised deep learning.
• To overcome the limitation of requiring matched reference data in existing deep learning approaches.
• To improve the accuracy and efficiency of TOF-MRA image reconstruction.

Main Methods:

• A two-stage unsupervised deep learning framework was proposed, extending CycleGAN from optimal transport theory.
• The first stage involved a multi-coil reconstruction network in the square-root of sum of squares (SSoS) domain.
• The second stage employed a multi-planar refinement network with a double-headed projection discriminator for blood flow characteristics.

Main Results:

• The unsupervised approach achieved high-quality parallel image reconstruction.
• The method successfully learned characteristics of highly-activated blood flow.
• Experimental results demonstrated superior performance compared to compressed sensing (CS) methods.
• The proposed method provided comparable or better results than supervised learning approaches.

Conclusions:

• The novel two-stage unsupervised deep learning method effectively reconstructs undersampled TOF-MRA.
• This approach eliminates the need for matched reference data, simplifying the training process.
• The method offers a promising alternative for high-quality, accelerated TOF-MRA acquisition and reconstruction.