Diffusion models for ultrasound imaging of sound-speed – Zezheng Zhang

Speed-of-sound is a promising quantitative imaging biomarker for tissue characterization and for enabling broader diagnostic applications of ultrasound (US) imaging. It is also a critical parameter for other US imaging modalities.
Therefore, high-quality speed-of-sound (SoS) map reconstruction is a critical yet challenging task.

Traditional algebraic algorithms are computationally intensive, rely on hand-crafted priors with carefully tuned regularization weights, and may produce over-smoothed results. In contrast, standard deep learning methods often suffer from overfitting to the training data and fail to generalize across domains, leading to blurry or physically inconsistent outputs when applied to real data, due to their ill-posedness in many scenarios.

In this talk, I will present a framework founded on a conditional diffusion model, which generates high-fidelity SoS maps by leveraging a learned data-driven prior to guide an iterative denoising process, while enforcing physical plausibility at each step through a data consistency optimiza- tion. The proposed framework demonstrated an improvement in reconstruction accuracy for SoS imaging.