MRI reconstruction using deep Bayesian estimation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Magnetic Resonance Imaging (MRI) reconstruction from incomplete k-space data is crucial for efficient imaging.
• Existing deep learning methods for MRI reconstruction face challenges in preserving image details and mitigating artifacts.

Purpose of the Study:

• To develop a novel deep learning-based Bayesian estimation framework for MRI reconstruction.
• To enhance image quality by improving detail preservation and reducing aliasing artifacts.

Main Methods:

• Modeled MRI reconstruction using Bayes's theorem and a generative network as an image prior, inspired by PixelCNN++.
• Employed maximum a posteriori estimation with stochastic backpropagation and projected subgradient methods for constraint enforcement.
• Utilized the likelihood of the prior as the training loss and objective function for reconstruction.

Main Results:

• The proposed Bayesian estimation method significantly improved MRI reconstruction performance.
• Demonstrated superior image detail preservation and reduction of aliasing artifacts compared to GRAPPA, ESPRiT, MODL, and VN.
• Achieved over 3 dB peak signal-to-noise ratio improvement in compressed sensing and parallel imaging reconstructions.

Conclusions:

• Bayesian estimation offers a significant performance improvement over conventional sparsity priors in compressed sensing MRI.
• The developed reconstruction framework is generalizable across various MRI reconstruction scenarios.
• This deep learning approach advances MRI reconstruction quality and efficiency.