Full-Wave Inversion for Speed-of-Sound Imaging in Ultrasound Tomography via Self-Supervised Learning

Full-waveform inversion (FWI) based ultrasound computed tomography (USCT) is a promising technique for quantitative imaging of soft tissues, which is particularly useful for breast cancer screening. However, its image reconstruction problem for speed-of-sound (SoS) imaging is a highly ill-conditioned inverse problem and suffers from severe cycle skipping, which often leads to instable reconstruction and image artifacts. In this study, we propose a self-supervised deep learning-based FWI framework to enhance the stability and accuracy of SoS reconstruction. In this framework, the slowness field is modeled as a continuous function using a coordinate-based neural network. The network is trained in a self-supervised manner by minimizing the mismatch between the predicted and observed pressure fields. Simulation experiments using the k-Wave toolbox with a 512-element ring array on human breast phantoms demonstrate that the proposed method significantly outperforms the traditional conjugate gradient approach, effectively reducing artifacts and mitigating cycle skipping. These results highlight the potential of self-supervised deep learning-based FWI as a robust and accurate method for SoS imaging in USCT.