JotlasNet: Joint tensor low-rank and attention-based sparse unrolling network for accelerating dynamic MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Dynamic MRI reconstruction faces challenges with existing methods that overlook tensor characteristics and use global sparse constraints.
• Current deep unrolling networks often have complex structures, limiting flexibility and efficiency.

Purpose of the Study:

• To introduce JotlasNet, a novel deep unrolling network for dynamic MRI reconstruction.
• To leverage tensor low-rank and attention-based sparse priors for improved reconstruction accuracy and flexibility.

Main Methods:

• Utilizing tensor low-rank prior to capture high-dimensional data correlations.
• Employing convolutional neural networks for adaptive learning of low-rank and sparse domains.
• Introducing an attention-based soft thresholding operator for channel-specific sparse constraints.

Main Results:

• JotlasNet demonstrates superior performance in dynamic MRI reconstruction.
• The network effectively exploits tensor characteristics and adaptive sparse priors.
• Experiments on OCMR and CMRxRecon datasets validate the proposed method's efficacy.

Conclusions:

• JotlasNet offers a significant advancement in dynamic MRI reconstruction.
• The proposed tensor low-rank and attention-based sparse priors enhance reconstruction quality.
• The network's simple, parallel structure contributes to its efficiency.