Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Diffusion as Sound Propagation: Physics-inspired Model for Ultrasound Image Generation

Marina Domínguez, Yordanka Velikova, Nassir Navab, Mohammad Farid Azampour

TL;DR

This paper tackles the challenge of limited labeled ultrasound data by proposing a physics-informed diffusion framework tailored to ultrasound imaging. It introduces B-Maps, a noise scheduler that mimics depth-dependent attenuation, and integrates them into forward and reverse diffusion processes within Guided-Diffusion and Semantic Diffusion Models. The approach is evaluated across SegThy, CAMUS, and Liver datasets, showing consistent improvements in FID, LPIPS, and PSNR (with SSIM largely maintained) compared to baselines. The findings demonstrate that incorporating ultrasound physics into diffusion-based data generation yields more realistic images, enhancing data augmentation and potential downstream learning tasks in medical imaging.

Abstract

Deep learning (DL) methods typically require large datasets to effectively learn data distributions. However, in the medical field, data is often limited in quantity, and acquiring labeled data can be costly. To mitigate this data scarcity, data augmentation techniques are commonly employed. Among these techniques, generative models play a pivotal role in expanding datasets. However, when it comes to ultrasound (US) imaging, the authenticity of generated data often diminishes due to the oversight of ultrasound physics. We propose a novel approach to improve the quality of generated US images by introducing a physics-based diffusion model that is specifically designed for this image modality. The proposed model incorporates an US-specific scheduler scheme that mimics the natural behavior of sound wave propagation in ultrasound imaging. Our analysis demonstrates how the proposed method aids in modeling the attenuation dynamics in US imaging. We present both qualitative and quantitative results based on standard generative model metrics, showing that our proposed method results in overall more plausible images. Our code is available at https://github.com/marinadominguez/diffusion-for-us-images

Diffusion as Sound Propagation: Physics-inspired Model for Ultrasound Image Generation

TL;DR

This paper tackles the challenge of limited labeled ultrasound data by proposing a physics-informed diffusion framework tailored to ultrasound imaging. It introduces B-Maps, a noise scheduler that mimics depth-dependent attenuation, and integrates them into forward and reverse diffusion processes within Guided-Diffusion and Semantic Diffusion Models. The approach is evaluated across SegThy, CAMUS, and Liver datasets, showing consistent improvements in FID, LPIPS, and PSNR (with SSIM largely maintained) compared to baselines. The findings demonstrate that incorporating ultrasound physics into diffusion-based data generation yields more realistic images, enhancing data augmentation and potential downstream learning tasks in medical imaging.

Abstract

Deep learning (DL) methods typically require large datasets to effectively learn data distributions. However, in the medical field, data is often limited in quantity, and acquiring labeled data can be costly. To mitigate this data scarcity, data augmentation techniques are commonly employed. Among these techniques, generative models play a pivotal role in expanding datasets. However, when it comes to ultrasound (US) imaging, the authenticity of generated data often diminishes due to the oversight of ultrasound physics. We propose a novel approach to improve the quality of generated US images by introducing a physics-based diffusion model that is specifically designed for this image modality. The proposed model incorporates an US-specific scheduler scheme that mimics the natural behavior of sound wave propagation in ultrasound imaging. Our analysis demonstrates how the proposed method aids in modeling the attenuation dynamics in US imaging. We present both qualitative and quantitative results based on standard generative model metrics, showing that our proposed method results in overall more plausible images. Our code is available at https://github.com/marinadominguez/diffusion-for-us-images
Paper Structure (20 sections, 11 equations, 7 figures, 2 tables)

This paper contains 20 sections, 11 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Contributions
  3. Methodology
  4. Background
  5. Forward process.
  6. Reverse process.
  7. B-Maps Definition
  8. Denoising Diffusion Probabilistic Models with B-Maps
  9. Forward Process.
  10. Reverse Process.
  11. Experimental Setup
  12. Datasets
  13. SegThy
  14. CAMUS
  15. Liver
  16. ...and 5 more sections

Figures (7)

  • Figure 1: Evolution of B-maps across time-steps. In every timestep, the values in the B-Maps decrease top-to-bottom from 1 to a number, $\gamma$. As the timestep increases, $\gamma$ goes from 1 to $1 - \epsilon$, with $\epsilon$ being a small fixed value in the interval $(0,1)$.
  • Figure 2: Forward pass: Noise addition from bottom to top. Linearly-scheduled cone-shaped B-Maps on the top row and the visualization of the noising process of the US image in the bottom row. B-Maps are applied at each step, making the gaussian distribution converge earlier on the bottom than on the top.
  • Figure 3: Reverse Process: denoising the image. Initially focusing on the area near the probe, the model progresses to denoise the image toward the bottom, mimicking the way US images are traditionally generated.
  • Figure 4: Qualitative comparison: The top row displays the label maps used for SegThy and CAMUS datasets. For the liver dataset, no labels were used. The bottom row shows the US images generated with B-Maps (left) versus without B-Maps (right) for each dataset.
  • Figure 5: Liver Results: US images generated with B-Maps (top) exhibit enhanced contrast, especially in the upper regions, compared to those without (bottom).
  • ...and 2 more figures