Physics-Informed DeepMRI: k-Space Interpolation Meets Heat Diffusion.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Diffusion models show promise for MRI reconstruction but suffer from artifacts due to inherent randomness.
• Existing methods often struggle with controlled image generation, particularly in high-frequency regions of k-space data.

Purpose of the Study:

• To develop a deterministic MRI reconstruction method using diffusion models grounded in low-frequency k-space data.
• To improve the accuracy and reduce artifacts in MRI reconstruction by focusing on high-frequency data interpolation.

Main Methods:

• Established a relationship between high-frequency (HF) k-space data interpolation and the reverse heat diffusion process.
• Developed a diffusion model incorporating a physics-informed k-space interpolation model as a data fidelity term.
• Utilized publicly available datasets for experimental validation and generalization assessment.

Main Results:

• The proposed diffusion model significantly outperforms traditional and deep learning-based k-space interpolation methods.
• Superior performance was observed particularly in reconstructing high-frequency k-space data.
• The model demonstrated robust generalization performance across various out-of-distribution datasets.

Conclusions:

• The deterministic, physics-informed diffusion model offers a significant advancement in MRI reconstruction accuracy and artifact reduction.
• This approach provides a fundamental framework for designing more controlled and effective diffusion-based image generation models.
• The method shows promise for improving the quality of reconstructed MRI images, especially in challenging high-frequency regions.