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

Accurately imaging the spatial distribution of longitudinal sound speed has a profound impact on image quality and the diagnostic value of ultrasound. Knowledge of sound speed distribution allows effective aberration correction to improve the image quality. Sound speed imaging also provides a new mechanism of imaging contrast that will facilitate disease diagnosis. However, sound speed imaging is challenging in the pulse-echo mode. Deep learning (DL) is a promising approach for pulse-echo speed of sound imaging, which may yield more accurate results than pure physics-based approaches. However, it often requires a large amount of training data and has limitations in model generalizability. To address these issues, we developed a physics-guided DL approach to image the spatial distribution of sound speed using pulse-echo data acquired from multiple angles. Instead of using the raw echo signals as the input to the DL model, we transformed the echo signals into time-shift maps based on local phase shifts between beamformed images. Using this approach, our model was shown to be generalizable when trained and tested using different scan settings in simulation studies. Furthermore, the simulation-trained model successfully reconstructed the sound speed maps of phantoms using experimental data. Compared with a state-of-the-art physics-based inversion approach, our approach reduced median root mean squared error (RMSE) by 26% and improved contrast-to-noise ratio (CNR) by 52% in the phantom experiment. These results demonstrated the accuracy and robustness of our approach.