High-precision wavefield simulation and deep learning-based sound speed reconstruction for transcranial ultrasound imaging.

Abstract

Transcranial ultrasound imaging plays an important role in the diagnosis of brain diseases and the monitoring of brain function. However, the quality of transcranial imaging is often impaired by the intricate acoustic properties of the skull. Accurate reconstruction of the skull's speed of sound (SoS) is critical for effective phase correction and enhanced image quality. In this study, we propose a transcranial SoS local reconstruction framework that integrates high-fidelity 2D numerical simulation with deep learning inversion. A custom wavefield simulation algorithm is developed to generate training datasets that can model spatially varying velocity and attenuation distributions. In the learning framework, we propose WAM-Net, which incorporates a Wavefront Attention Module (WAM) and a gradient-regularized loss function to reconstruct the skull's SoS accurately. In numerical simulations, the proposed WAM-Net method significantly improves reconstruction speed compared to full-waveform inversion (FWI), and reduces the SoS reconstruction error by 63.52% compared to AutoSoS. In skull-mimicking phantom experiments, the method demonstrates reliable SoS reconstruction across various inclinations and structural designs, with an average Mean Absolute Error (MAE) of 13.4844 m/s in Al₂O₃ phantom and a MAE of 31.3804 m/s in PMMA phantom. In the in-vivo experiments on a crab-eating macaque, the constructed SoS map effectively distinguishes between dense bone and porous bone in anatomically complex regions. These results indicate that the method provides an effective solution for real-time transcranial aberration correction, with high structural fidelity and robustness in heterogeneous cranial environments.