Generalizable Deep Learning for Pulse-echo Speed of Sound Imaging via Time-shift Maps

Accurate speed of sound (SoS) imaging is important for improving the diagnostic capability and image quality of ultrasound systems. However, achieving accurate SoS imaging in the pulse-echo mode remains challenging. While deep learning (DL) approaches have shown promise in SoS reconstruction from simulated echo data, they often suffer from significant performance degradation when transitioning from simulation to experimental data, highlighting a gap in generalizability. To address this gap, we propose a DL framework that utilizes time-shift maps as the input for SoS reconstruction. The time-shift maps are obtained from raw echo data through customized beamforming and phase-shift tracking. Compared with raw echo data, the proposed time-shift measurement is more directly linked to SoS, based on the physical principle that SoS variation causes the shifts in time of flight. Simulation studies demonstrate that our method performs reliably across varying conditions, reducing the influence of pulse settings and medium properties. Experiments with tissue-mimicking phantoms show that the simulation-trained model successfully generalizes to real-world data.