Data-Driven full waveform inversion for ultrasonic bone quantitative imaging

Full waveform inversion (FWI) has potential for quantitative ultrasound bone imaging, which can provide detailed estimation of bone internal structure. The existing ultrasound FWI methods suffer from cycle skipping and local minima and high computational costs, which limits their clinical application potential. In recent years, data-driven method for solving the problem of inversion has become a very novel approach, but we haven't seen it in the application of ultrasonic bone imaging studies. Herein, we develop an improved dual-encoder-based Unet with high-frequency feature enhancement (DEFE-Unet) for ultrasonic bone quantitative imaging to obtain more detailed information by inserting high-frequency feature extractors related to bone microstructure into the network. This method can obtain more detailed and fine results than FWI and avoid the extremely high computational cost of FWI dual-parameter modeling. We designed two types of bone model datasets based on Micro_CT images to train and test the proposed network architecture. Experimental results show that the DEFE-Unet algorithm proposed in this paper obtains competitive results with less computational cost compared to the FWI algorithm. Compared with other baseline machine learning architectures, such as Unet and attention-Unet, the proposed algorithm achieves significant performance improvement with a small computational cost increment. The results show that the DEFE-Unet has better potential in clinical detection of bone disease.