WKGM: weighted k-space generative model for parallel imaging reconstruction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Deep learning has accelerated Magnetic Resonance Imaging (MRI) through parallel imaging (PI).
• Existing PI methods often lack robustness and flexibility.
• Calibrationless reconstruction remains a challenge in accelerated MRI.

Purpose of the Study:

• To develop a robust and flexible deep learning model for calibrationless parallel imaging (PI) in MRI.
• To explore k-space domain learning using generative modeling for improved MRI reconstruction.
• To enhance the efficiency and applicability of accelerated MRI techniques.

Main Methods:

• Proposed the weighted k-space generative model (WKGM), a generalized k-space domain model.
• Incorporated k-space weighting and high-dimensional space augmentation for score-based generative model training.
• Designed WKGM for synergistic combination with traditional k-space PI models, leveraging multi-coil data correlations.

Main Results:

• Achieved good and robust MRI reconstructions using the WKGM.
• Demonstrated flexibility by combining WKGM with existing PI models for calibrationless reconstruction.
• Experimental results showed state-of-the-art performance even with a limited training dataset (500 images) and varying sampling patterns.

Conclusions:

• WKGM offers a robust and flexible solution for calibrationless parallel imaging in MRI.
• The model effectively learns k-space generative priors for high-quality image reconstruction.
• WKGM represents a significant advancement in accelerating MRI scans while maintaining image fidelity.