Feature-Space Guided Diffusion for Realistic Ultrasound Image Synthesis

2026-07-13

AI Summary

This work addresses the visual distortion in ultrasound images generated by existing conditional diffusion models, which, despite anatomically plausible structures, lack key imaging characteristics of real B-mode ultrasound—such as speckle noise, tissue contrast, and acoustic attenuation. To bridge this domain gap without additional training, the authors propose a novel inference-time feature-space conditional guidance (FSCG) strategy that leverages a frozen foundational ultrasound model to steer the generation process. FSCG integrates local k-nearest neighbor feature correction with energy-based evaluation and selection among multiple candidate samples, preserving the prescribed anatomical structure while aligning the output with the real ultrasound data distribution. Experiments across three datasets demonstrate that FSCG reduces FID64 and FID192 by 56% and 57%, respectively, and decreases nearest-neighbor feature distance by 47%, substantially outperforming current inference-time guidance methods.

Abstract

Conditional diffusion models can generate anatomically plausible medical ultrasound (US) images, but anatomical plausibility alone does not ensure realistic B-mode appearance. Most US pipelines adapt standard generative architectures and condition them on anatomical masks, or use guidance mechanisms that reinforce the same anatomical signal. However, B-mode US images are shaped by acquisition-dependent properties such as speckle texture, tissue contrast, and attenuation. Using a frozen US foundation model, we show that standard conditional diffusion baselines remain separated from real images in representation space. In this work, we propose Feature-Space Candidate Guidance (FSCG), a training-free sampling strategy to reduce this gap. At sampling time, FSCG applies local k-NN feature correction and selects the best of multiple stochastic candidates according to their feature-space energy. In this way, the mask defines the anatomy, while FSCG steers samples toward the real US domain. Across three different datasets, FSCG reduces average FID64 by 56\%, FID192 by 57\%, and nearest-neighbour feature distance by 47\% over standard conditional diffusion sampling, outperforming alternative inference-time guidance baselines. The results suggest that domain-aware feature representations can reveal and reduce realism gaps in medical diffusion synthesis without retraining the generator. Our code is available at https://github.com/marinadominguez/FSCG.

Problem

• ultrasound image synthesis
• realism gap
• feature-space representation
• conditional diffusion models
• B-mode appearance

Innovation

• Feature-Space Guidance
• Ultrasound Image Synthesis
• Conditional Diffusion Models
• Training-Free Sampling
• Medical Image Generation