Abdominal Sound Speed Estimation through Machine Learning

In clinical ultrasound, the assumption of a uniform sound speed (SoS) of 1540 m/s can lead to errors, especially in areas with bone or bowel gas. Through-transmission ultrasound can produce quantitative SoS images, but waveform inversion (WI) methods used for refining SoS models rely heavily on an accurate initial model. This work presents a customized U-Net neural network trained on synthetic ultrasound data to provide more accurate initial SoS models for abdominal cross-sections, improving WI convergence. Forward-simulated acoustic data were generated using a 64-element semicircular array, and the neural network estimated SoS maps from radial B-mode lines. The network's performance was evaluated using Dice coefficient, SSIM, PSNR, and SoS difference metrics, showing superior results compared to a uniform 1540 m/s model. Hyperparameter tuning demonstrated that deeper networks with dropout rates of 50% achieved optimal performance. Overall, our findings suggest that neural networks can enhance WI processes by offering more accurate SoS models than traditional homogeneous assumptions.