A Conditional Normalizing Flow for Accelerated Multi-Coil MR Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Accelerated magnetic resonance (MR) imaging aims to reduce scan times by acquiring data below the Nyquist rate.
• This undersampling results in an ill-posed inverse problem with multiple possible solutions.
• Current deep learning methods often produce a single solution, limiting downstream inference.

Purpose of the Study:

• To develop a novel deep learning approach for accelerated MR imaging that samples from the posterior distribution.
• To provide more comprehensive information for downstream inference tasks compared to single-solution methods.
• To improve the speed and accuracy of accelerated MR image reconstruction.

Main Methods:

• Designed a novel conditional normalizing flow (CNF) model.
• The CNF infers the signal component within the measurement operator's nullspace.
• This inferred component is combined with measured data to reconstruct complete MR images.

Main Results:

• Demonstrated fast inference times on the fastMRI brain and knee datasets.
• Achieved accuracy surpassing recent posterior sampling techniques for MR imaging.
• The CNF approach effectively reconstructs complete MR images from undersampled data.

Conclusions:

• The proposed conditional normalizing flow (CNF) offers a powerful method for accelerated MR imaging.
• Sampling from the posterior distribution provides richer information than single-solution approaches.
• This technique advances the field of fast and accurate medical image acquisition.