MRI Super-Resolution With Partial Diffusion Models

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Diffusion models excel at image generation tasks like super-resolution.
• High computational demands limit their practical application due to numerous denoising steps.

Purpose of the Study:

• To introduce Partial Diffusion Models (PDMs) for accelerated magnetic resonance imaging (MRI) super-resolution.
• To reduce the computational cost of diffusion models in MRI super-resolution.

Main Methods:

• Proposed PDMs leverage the convergence of low- and high-resolution image latents at certain noise levels.
• Introduced 'latent alignment' to mitigate approximation errors by interpolating latents.
• Developed a new diffusion trajectory from low- to high-resolution images.

Main Results:

• PDMs achieve competitive super-resolution quality on MRI data.
• Significantly fewer denoising steps are required compared to standard diffusion models.
• PDMs can be combined with existing acceleration techniques for further efficiency gains.

Conclusions:

• PDMs offer an efficient solution for MRI super-resolution.
• The method significantly reduces computational complexity without sacrificing image quality.
• This work paves the way for faster and more accessible MRI super-resolution techniques.