Deep-Learning Inversion for Bone Ultrasound Tomography

As a noninvasive medical imaging modality, ultrasound offers the advantages of safety, convenience, and affordability. However, bone imaging has long been a challenge due to the high acoustic impedance. Inversion algorithms have shown promise in advancing ultrasound computed tomography (UCT) as a viable technique for bone imaging. Traditional physics-driven inversion algorithms, such as full-waveform inversion (FWI), are computationally intensive and tend to be trapped in local minima. In this study, we propose a deep-learning (DL) inversion method to achieve fast bone imaging. This data-driven approach establishes a mapping from ultrasound data to bone sound speed (SoS) maps. Moreover, a physics-informed prior knowledge enhancement module is designed to extract and fuse travel-time and low-frequency information from the data. This enables the network to focus on key features based on physical principles and improves its generalization capability under waveform-domain variations. In simulation tests, the DL inversion can reconstruct a cortical bone image within two seconds and achieves a mean structural similarity index measure (SSIM) of 0.9823. Although trained on data from a specific source, the proposed network remains robust to waveform -domain mismatches, such as changes in pulselength, center frequency, measurement noise, and propagation effects introduced by soft tissues. In laboratory validation, despite being trained on simulated datasets, the network is able to reconstruct bone phantom images using experimental data, with an SSIM of 0.9528. This demonstrates the robustness of the prior-informed DL inversion method for bone imaging, which offers a rapid, accurate, and adaptable solution that overcomes the limitations of traditional inversion methods.