Highly accelerated MRI via implicit neural representation guided posterior sampling of diffusion models.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Accelerating magnetic resonance (MR) imaging acquisition is crucial for reducing scan times and improving patient comfort.
• Traditional methods for reconstructing MR images from under-sampled k-space data often struggle with accuracy and stability due to insufficient data consistency.
• Diffusion models show promise for MR image reconstruction, but existing posterior sampling techniques lack robust guidance.

Purpose of the Study:

• To develop a novel posterior sampler for diffusion models that integrates implicit neural representation (INR) for enhanced MR image reconstruction.
• To improve the accuracy, generalizability, and stability of MR image reconstruction, particularly under high acceleration factors.
• To create a generalizable framework for solving inverse problems in medical imaging.

Main Methods:

• Introduced DiffINR, a novel posterior sampler for diffusion models that utilizes implicit neural representation (INR).
• Integrated the diffusion prior distribution and the MR physical model within the INR component to ensure data fidelity.
• Evaluated DiffINR on in-distribution datasets with varying acceleration factors (up to R=12) and across different tissue contrasts and anatomical structures.

Main Results:

• DiffINR demonstrated superior performance with remarkable accuracy in MR image reconstruction, even at high acceleration factors.
• The method exhibited excellent generalizability across diverse tissue contrasts and anatomical structures, with low reconstruction uncertainty.
• Achieved significant improvements in accuracy, generalizability, and stability compared to traditional posterior sampling methods.

Conclusions:

• DiffINR significantly enhances the accuracy, generalizability, and stability of MR image reconstruction, enabling faster scan times.
• The proposed INR-based posterior sampler offers a promising approach for accelerating MR acquisition.
• The DiffINR framework shows potential for broader application in solving inverse problems across various medical imaging modalities.