Learning Data Consistency and its Application to Dynamic MR Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Magnetic Resonance (MR) image reconstruction from undersampled k-space data is crucial for faster scans.
• Current deep learning (DL) methods often enforce explicit data consistency, neglecting real noise distributions.
• This can limit the performance of MR reconstruction algorithms.

Purpose of the Study:

• To develop a novel DL-based approach for MR image reconstruction that implicitly learns data consistency.
• To address the limitations of existing methods by accounting for actual system noise distributions.
• To improve the quantitative and qualitative performance of MR image reconstruction.

Main Methods:

• Proposed a new DL-based approach termed Learned DC.
• Learned DC implicitly learns data consistency within deep networks, aligning with system noise probability distributions.
• Data consistency and image prior are embedded within the network weights for an implicit learning process.

Main Results:

• Evaluated Learned DC on highly undersampled dynamic cardiac cine and rectum data.
• Achieved up to 24-fold acceleration in cardiac cine imaging.
• Demonstrated superior quantitative and qualitative performance compared to state-of-the-art methods.

Conclusions:

• Learned DC offers a more effective approach to MR image reconstruction by implicitly learning data consistency.
• The method shows significant potential for accelerating dynamic MR imaging while maintaining high image quality.
• This implicit learning strategy represents a advancement in DL-based MR reconstruction.