Bayesian reconstruction of magnetic resonance images using Gaussian processes

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Applied Mathematics

Background:

• Accelerating magnetic resonance imaging (MRI) acquisition is crucial for clinical applications.
• Existing methods like parallel imaging, compressed sensing, and deep learning have limitations.
• Developing novel reconstruction techniques is essential for faster, high-quality MRI.

Purpose of the Study:

• To propose and demonstrate a Bayesian method for optimizing MRI k-space sampling and reconstruction.
• To leverage statistical image libraries and Gaussian processes for efficient data acquisition.
• To validate the method's performance and transferability across different datasets.

Main Methods:

• Computed a multivariate normal distribution using Gaussian processes on T1-weighted brain images.
• Combined image libraries with physics-informed functions to retain meaningful k-space correlations.
• Employed Bayesian optimization to select optimal, practical ring-shaped k-space subsampling paths.
• Developed a generalized sampling path for novel image reconstruction.

Main Results:

• Achieved 96.3% structural similarity and <0.003 normalized mean squared error with only 12.5% k-space data.
• Demonstrated superior performance compared to existing reconstruction methods.
• Successfully applied the reconstruction to pathological data (stroke identification) without retraining.
• Showcased the model's inherent transferability from healthy to pathological brain images.

Conclusions:

• The proposed Bayesian method enables highly efficient MRI acquisition and reconstruction.
• This approach significantly reduces scan time while preserving diagnostic image quality.
• The method's transferability suggests broad applicability in clinical MRI, including for disease detection.