High-Frequency Space Diffusion Model for Accelerated MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Diffusion models using continuous stochastic differential equations (SDEs) excel in image generation and can solve inverse problems in magnetic resonance (MR) reconstruction.
• Current diffusion models applied to MR reconstruction struggle with fully sampled low-frequency k-space data, leading to reconstruction uncertainty and slow convergence.
• Existing methods require many iterations, making MR image reconstruction time-consuming.

Purpose of the Study:

• To develop a novel SDE tailored for MR reconstruction that addresses limitations of existing diffusion models.
• To improve reconstruction accuracy and stability in fast MR imaging.
• To accelerate the MR image reconstruction process.

Main Methods:

• Proposing a novel SDE with the diffusion process in high-frequency space (HFS-SDE) for MR reconstruction.
• Ensuring determinism in fully sampled low-frequency regions and accelerating reverse diffusion sampling.
• Utilizing the publicly available fastMRI dataset for experimental validation.

Main Results:

• The HFS-SDE method demonstrates superior reconstruction accuracy and stability compared to traditional parallel imaging, supervised deep learning, and existing diffusion models.
• Fast convergence properties of the HFS-SDE method are validated both theoretically and experimentally.
• The proposed method effectively handles the reconstruction of low-frequency regions in k-space data.

Conclusions:

• The HFS-SDE approach offers a significant advancement in MR image reconstruction, overcoming key challenges of existing diffusion models.
• This method provides a more accurate, stable, and faster solution for fast MR imaging.
• The developed technique has the potential to improve clinical workflow efficiency in MR imaging.