core

Unsupervised Medical Image Translation with Adversarial Diffusion Models

Muzaffer Özbey, Onat Dalmaz, Salman UH Dar · Bilkent University · 2022-07 · arXiv:2207.08208

Why it matters

Translates between imaging modalities (multi-contrast MRI, MRI to CT) with an adversarial diffusion model trained without paired data, using cycle consistency. The argument made is diffusion versus GAN fidelity, not a claim about which space to work in.

What this paper does

read: full text

Before this

Medical image translation between modalities relied on conditional GANs mapping source to target in one shot, but GANs characterize the target distribution only implicitly, with no likelihood to evaluate, and are prone to discriminator collapse and mode collapse that limit sample diversity and fidelity.

The problem

Standard diffusion models offer better diversity and an explicit likelihood-based objective, but need hundreds of sequential denoising steps, too slow for practical translation, and prior work had no way to train such a model without paired source-target images, which most cross-modality medical datasets lack.

The idea

Combine diffusion with adversarial projections so a small number of large denoising steps can still be sampled accurately, and wrap the diffusive generator in a cycle-consistent, non-diffusive pair of modules so the whole system trains on unpaired data.

How it works

SynDiff generates directly in pixel space, with no learned latent involved anywhere. The stated reason for large steps is speed, since standard diffusion needs hundreds of steps and SynDiff instead takes a small number of large steps (step size 250, four steps total) for efficiency, but this breaks the Gaussian-transition assumption ordinary diffusion sampling relies on, so a dedicated adversarial projector is trained to model the resulting non-Gaussian transition probability directly. The paper never discusses using a VAE or other latent space; it gives no reason to prefer or avoid one, it simply never considers the option.

Evidence

Evaluation is distribution and image-quality metrics only, PSNR and SSIM against a paired ground-truth target image, with no downstream clinical task or reader study reported. On multi-contrast MRI translation (IXI, T2 to T1) SynDiff reaches 30.13 dB PSNR and 94.61% SSIM against cGAN's 29.22 dB and 93.46%, and DDPM's 24.93 dB and 89.49%. On MRI-to-CT translation (pelvic dataset) SynDiff reaches 26.86 dB PSNR and 87.94% SSIM against cGAN's 25.07 dB and UNIT's 26.10 dB. The paper frames the 3.56 dB PSNR and 3.80-point SSIM gain over DDPM as evidence that adversarial projection recovers the quality standard diffusion loses when forced to take large steps.

Limitations

The paper concedes the method is built for one-to-one modality translation and would need architectural changes to use more than one source modality per prediction, and suggests transformer-based attention and larger training sets as directions likely to improve generalization further, rather than presenting the current result as final.

Why it matters

The contribution is about sampling efficiency and distributional fidelity for diffusion versus GANs, not about which representation space to generate in. It sits in the pixel-space part of the medical atlas as a data point that even fast, high-quality diffusion translation here was validated only against image-quality metrics, not clinical downstream performance.

Abstract, in the authors' own words

Imputation of missing images via source-to-target modality translation can improve diversity in medical imaging protocols. A pervasive approach for synthesizing target images involves one-shot mapping through generative adversarial networks (GAN). Yet, GAN models that implicitly characterize the image distribution can suffer from limited sample fidelity. Here, we propose a novel method based on adversarial diffusion modeling, SynDiff, for improved performance in medical image translation. To capture a direct correlate of the image distribution, SynDiff leverages a conditional diffusion process that progressively maps noise and source images onto the target image. For fast and accurate image sampling during inference, large diffusion steps are taken with adversarial projections in the reverse diffusion direction. To enable training on unpaired datasets, a cycle-consistent architecture is devised with coupled diffusive and non-diffusive modules that bilaterally translate between two modalities. Extensive assessments are reported on the utility of SynDiff against competing GAN and diffusion models in multi-contrast MRI and MRI-CT translation. Our demonstrations indicate that SynDiff offers quantitatively and qualitatively superior performance against competing baselines.

Method note — the shared flow-matching interpolation

Every flow-matching / rectified-flow paper in this atlas trains toward a straight-line path between a noise sample x₀ and a data sample x₁:

x_t = (1-t)\,x_0 + t\,x_1, \qquad v_\theta(x_t, t) \approx x_1 - x_0

Builds on

Nothing recorded yet.

Built on by

Nothing recorded yet.