Untrained Neural Network-Based Full-Waveform Inversion for Breast Sound Speed Imaging in Ultrasound Computed Tomography

Quantitative imaging represents a pivotal advantage of ultrasound computed tomography (USCT), significantly distinguishing it from conventional ultrasound imaging techniques. Especially, sound speed imaging in USTC has been documented to play a crucial role in the clinical diagnosis of breast cancer. Due to the high demand for resolution in medical applications, full-waveform inversion (FWI) has gradually become a research hotspot compared to time-of-flight methods due to its potential for high-resolution imaging. However, the main challenge of FWI is its sensitivity to initial values due to its inherent nonconvex and nonlinear properties. An inappropriate initial value can lead it to fall into a local optimal solution, which ultimately produces a sound speed image with no physiological significance, a phenomenon also known as “cycle-skipping,” severely limiting the application of the method to USCT. In this article, we propose a method based on an untrained neural network (UNN), which is embedded into the framework of the conventional iteration-based FWI as a sound speed image generator. The implicit regularization provided by this method can effectively alleviate the ill-posed nature of FWI. Significantly, the proposed UNN method allows for unsupervised training, which is essential in medical imaging due to the lack of ground-truth paired data. Evaluations on numerical simulations, phantom experiments, and in vivo breast experiments demonstrate the effectiveness of the proposed UNN in enhancing imaging robustness and reducing image artifacts compared to the traditional FWI. To the best of authors’ knowledge, this study represents the first to investigate the UNN-based FWI for USCT breast sound speed imaging and validate it on in vivo breast experiments.