Latent-k-space of refinement diffusion model for accelerated MRI reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Diffusion models (DM) show promise for magnetic resonance imaging (MRI) reconstruction.
• Current DM methods are computationally expensive due to pixel-level modeling and numerous iterations.
• Image-domain processing in existing methods can introduce secondary artifacts.

Purpose of the Study:

• To develop a novel diffusion model for accelerated MRI reconstruction.
• To reduce computational costs and artifacts associated with current DM-based MRI reconstruction.
• To improve the efficiency and quality of MRI image reconstruction.

Main Methods:

• Proposed a latent-k-space refinement diffusion model (LRDM) for MRI reconstruction.
• Encoded k-space data into a compact latent space for primary feature capture.
• Applied DM in the low-dimensional latent-k-space for prior generation and a secondary DM for high-frequency refinement.

Main Results:

• The LRDM requires only 4 iterations for accurate prior generation due to latent-space diffusion.
• The method significantly reduces MRI reconstruction time compared to conventional DM approaches.
• Achieved comparable image reconstruction quality to existing DM-based methods.

Conclusions:

• The LRDM offers a computationally efficient and effective solution for MRI reconstruction.
• Latent-space diffusion combined with high-frequency refinement enhances reconstruction speed and quality.
• This approach addresses key limitations of current diffusion model applications in MRI.