Zero-Shot Dynamic MRI Reconstruction With Global-to-Local Diffusion Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Diffusion Models

Background:

• Diffusion models show promise for MRI reconstruction but face challenges in dynamic MRI due to extensive training data needs.
• Acquiring fully sampled dynamic MRI data is difficult due to spatiotemporal complexity and high costs.

Purpose of the Study:

• To propose a zero-shot learning framework for accurate dynamic MRI reconstruction from undersampled k-space data.
• To address the limitations of current methods by eliminating the need for fully sampled training data.

Main Methods:

• A time-interleaved acquisition scheme merges undersampled k-space data from adjacent frames to create pseudo fully encoded reference data.
• A two-stage refinement strategy within the diffusion process learns global-to-local priors for effective data distribution capture.
• The framework enables zero-shot reconstruction, directly learning from undersampled data.

Main Results:

• The proposed method achieves accurate dynamic MR image reconstruction from undersampled k-space data.
• It effectively reduces noise and preserves image details.
• Reconstruction quality is comparable to supervised learning approaches.

Conclusions:

• The zero-shot learning framework successfully overcomes the need for fully sampled training data in dynamic MRI reconstruction.
• The method demonstrates robust performance in noise reduction and detail preservation.
• This approach offers a viable solution for efficient and high-quality dynamic MRI reconstruction.