High-resolution MRI synthesis using a data-driven framework with denoising diffusion probabilistic modeling

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Deep Learning for Image Reconstruction

Background:

• High-resolution magnetic resonance imaging (MRI) is crucial for accurate lesion diagnosis and delineation.
• Current limitations in gradient power and hardware restrict MRI resolution (sub-1 mm slices).
• Long MRI scan times are clinically unacceptable, hindering the acquisition of high-resolution images.

Purpose of the Study:

• To develop a framework for generating high-resolution MRI from low-resolution images using diffusion probabilistic deep learning.
• To improve the uncertainty and quality of denoising diffusion probabilistic models (DDPM) for MRI super-resolution.
• To overcome the limitations of conventional methods in capturing complex, high-dimensional image data.

Main Methods:

• A diffusion probabilistic deep learning framework utilizing denoising diffusion probabilistic models (DDPM) was developed.
• The forward process involved systematically adding Gaussian noise to low-resolution MRI images.
• The reverse process trained a U-Net model to denoise images and generate high-resolution outputs conditioned on low-resolution counterparts, tested on prostate and brain MRI datasets (BraTS2020).

Main Results:

• The proposed DDPM framework improved noise quality by 12.8% for prostate MRI, outperforming Bicubic (4.4%) and CGAN (5.7%).
• Signal-to-noise ratios were enhanced by 11.7% with DDPM, exceeding Bicubic (9.8%) and CGAN (8.1%).
• For BraTS2020 data, DDPM achieved a peak signal-to-noise ratio enhancement of 9.1%, with high multi-scale structural similarity (0.970 ± 0.019).

Conclusions:

• The developed deep learning-based diffusion probabilistic framework effectively enhances MRI resolution.
• This approach enables the acquisition of high-resolution MRI images without extending scan times, potentially improving clinical workflows.
• Future research will focus on prospectively validating the framework's efficacy across various clinical indications.