Low-rank tensor assisted K-space generative model for parallel imaging reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning, particularly generative models, shows promise in magnetic resonance imaging (MRI) but requires further enhancement.
• Score-based generative models' performance is sensitive to sample size and internal dimensions for gradient estimation.

Purpose of the Study:

• To introduce a novel parallel imaging reconstruction method, the low-rank tensor assisted k-space generative model (LR-KGM).
• To leverage low-rank information as high-dimensional prior information for generative models in MRI.

Main Methods:

• Constructing multi-channel MRI data into a large Hankel matrix to reduce training samples.
• Collapsing the Hankel matrix into a tensor for prior learning.
• Employing a low-rank rotation strategy during testing to constrain generative network output tensors.
• Alternating reconstruction between traditional generative and low-rank high-dimensional tensor iterations.

Main Results:

• The proposed LR-KGM method demonstrated superior performance compared to existing state-of-the-art techniques.
• The integration of low-rank tensor priors improved the accuracy and efficiency of MRI reconstruction.

Conclusions:

• The LR-KGM method offers a significant advancement in parallel imaging reconstruction for MRI.
• Transforming low-rank information into high-dimensional priors effectively enhances generative model performance in MRI applications.