Unsupervised Medical Image Translation With Adversarial Diffusion Models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical imaging protocols benefit from imputing missing data through source-to-target modality translation.
• Generative Adversarial Networks (GANs) are commonly used for synthesizing target images but can have limited sample fidelity.
• Existing methods struggle with the implicit characterization of image distributions.

Purpose of the Study:

• To introduce SynDiff, a novel adversarial diffusion modeling method for enhanced medical image translation.
• To improve sample fidelity and performance in synthesizing target medical images.
• To enable training on unpaired datasets for bilateral modality translation.

Main Methods:

• SynDiff utilizes a conditional diffusion process to map noise and source images onto the target image, capturing the image distribution directly.
• Large diffusion steps with adversarial projections in the reverse diffusion direction ensure fast and accurate inference.
• A cycle-consistent architecture with coupled diffusive and non-diffusive modules facilitates training on unpaired data.

Main Results:

• SynDiff demonstrates quantitatively and qualitatively superior performance compared to competing GAN and diffusion models.
• Extensive assessments were conducted on multi-contrast MRI and MRI-CT translation tasks.
• The proposed method achieves improved sample fidelity in medical image synthesis.

Conclusions:

• Adversarial diffusion modeling, as implemented in SynDiff, offers a powerful approach for medical image translation.
• SynDiff overcomes limitations of traditional GANs in sample fidelity for medical image synthesis.
• The method shows significant potential for improving diversity and imputation in medical imaging protocols.