Physics-based Speed of Sound Estimation in Ultrasound Tomography guided by Differentiable Physics

Full Waveform Inversion (FWI) remains the standard approach in Ultrasound Computed Tomography (UCT) for determining Speed of Sound (SoS) maps in tissue imaging, although it is computationally demanding. Supervised Deep Learning (SDL) techniques for learning the mapping from sensor data to ground-truth SoS maps have been investigated as a possible source of efficiency, but they do not generalize well to new datasets. Recent UCT research relies on introducing Differentiable Physics (DP) into the acoustic wave equation model. However, the inverse network design for sampling SoS maps remains unaltered, which is the motivating factor for this work. We propose an inverse network architecture based on wave equation physics, along with a DP-based forward simulator, to improve SoS estimation while incurring no additional expenses compared to SDL.