A Differentiable Physics Approach for Unsupervised Ultrasound Computed Tomography

Ultrasound Computed Tomography (USCT) is a widely used non-invasive medical imaging modality for obtaining high resolution Speed of Sound (SoS) maps of tissue structures in the body. To obtain these velocity maps, traditionally Full Waveform Inversion (FWI) has been used in solving the inverse of the corresponding acoustic wave propagation equation. More recently, Supervised Deep Learning (SDL) has been utilized for USCT, to alleviate the problems of FWI. However, DL requires large volumes of paired training data, and often suffers from generalization. In this work, we consider integrating Differentiable Physics (DP) with DL and propose a first of its kind, Unsupervised DP (UDP) approach to learn SoS for USCT, from only the transducer data recordings.