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

Marina Dominguez*, Yordanka Velikova*, Mohammad Farid Azampour, Nassir Navab

MICCAI 2024 (Oral Presentation) · arXiv:2407.05428

* equal contribution

arXiv Code (GitHub) BibTeX

Abstract

Deep learning workflows typically need large-scale labeled datasets, which are expensive to obtain for medical ultrasound. To mitigate the limited training data, we introduce a physics-aware diffusion process that mimics sound propagation via ultrasound-specific scheduler schemes. The model couples attenuation dynamics with B-map guidance to produce more realistic B-mode simulations and to better capture ultrasound-specific noise distributions. We demonstrate qualitative and quantitative improvements in plausibility and realism across multiple anatomical datasets.

Physics-guided diffusion

Forward diffusion with B-maps — Forward process: cone-shaped B-maps steer the noise addition to match ultrasound propagation.

Evolution of B-maps — B-map schedule decreases amplitudes from probe to depth, matching attenuation dynamics.

Reverse denoising process — Reverse process first clarifies regions near the probe before resolving deeper structures.

DiffUS teaser animation — Animated overview of the physics-inspired diffusion model.

Results

B-map guidance delivers sharper simulations across SegThy, CAMUS, and liver cohorts.

Citation

If you find this work useful, please cite:

@misc{dominguez2024diffus,
  title={Diffusion as Sound Propagation: Physics-inspired Model for Ultrasound Image Generation},
  author={Marina Dominguez and Yordanka Velikova and Mohammad Farid Azampour and Nassir Navab},
  year={2024},
  eprint={2407.05428},
  archivePrefix={arXiv},
  primaryClass={eess.IV}
}