Abstract
Ultrasound computed tomography provides speed-of-sound (SoS) images that complement conventional pulse–echo ultrasound, but estimating spatial acoustic structure from transmission data remains computationally demanding. We investigate a physics-informed learned reconstruction trained exclusively on synthetic SoS fields and Eikonal-calculated travel times. Diverse synthetic structures supply paired training data at scale, enabling learning without experimentally labeled SoS images. A supervised inverse, a latent image prior, and physics-conditioned corrections produce an image that can subsequently be refined through the Eikonal equation. Evaluation on held-out synthetic fields shows improved reconstruction relative to the supervised initial estimate, with further gains from post-inference physical refinement. Applied to ToFs extracted from an in vivo breast acquisition, the fixed learned models recover heterogeneous internal SoS structure, and refinement improves correspondence with an independent full-waveform-inversion reconstruction. Controlled analyses characterize the contributions of spatial conditioning and the dependence of reconstruction on timing, support, and numerical sampling. These findings demonstrate that synthetic structures can provide useful prior information for ultrasound travel-time reconstruction and establish the feasibility of combining rapid learned inference with image-specific physical refinement.