Synthetizing SWI from 3T to 7T by generative diffusion network for deep medullary veins visualization

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Ultrahigh-field (7 Tesla) susceptibility-weighted imaging (SWI) offers superior brain contrast and detail.
• However, 7T MRI scanners are costly and generate significant noise, impacting patient comfort.
• Existing deep learning methods, primarily Generative Adversarial Networks (GANs), face training challenges limiting their performance for SWI synthesis.

Purpose of the Study:

• To develop and evaluate a diffusion-based deep learning model for synthesizing 7T SWI images from 3T SWI images.
• To assess the clinical applicability of the proposed model for enhanced visualization of brain microvasculature.
• To overcome the limitations of GANs in synthesizing high-fidelity SWI images.

Main Methods:

• A conditional denoising diffusion probabilistic model (CDDPM) was developed for image synthesis.
• The CDDPM was trained to generate 7T SWI images using 3T SWI images as input.
• The model's performance was evaluated for its ability to synthesize high-field SWI characteristics.

Main Results:

• The diffusion-based CDDPM successfully synthesized high-field (7T) SWI images from low-field (3T) inputs.
• The synthesized images demonstrated potential for improved visualization of deep medullary veins.
• The model offers a promising alternative to traditional GAN-based synthesis methods.

Conclusions:

• The developed CDDPM provides a viable method for generating high-quality 7T SWI images from 3T data.
• This approach may offer a cost-effective and patient-friendly alternative to ultra-high field MRI.
• The diffusion model shows significant potential for clinical applications, particularly in visualizing deep medullary veins.