Abdominal Sound Speed Estimation Using Neural Networks Trained on Wave Propagation Physics

Abdominal diagnostic ultrasound imaging fails in more than 20% of cases due to the degrading effects of the body wall. Aberration estimation and correction is crucial, yet challenging, for improving medical diagnostics. To estimate the sound speed distribution throughout an imaged area for calculating spatially varying beamforming delays that account for tissue inhomogeneities, we have developed a neural network sound speed estimation approach that was trained from a simulated liver imaging dataset utilizing anatomically derived human body wall maps and full wave propagation physics through Fullwave 2. We created a dataset of abdominal radiofrequency (RF) data using segmentations from the Visible Human Project and trained the network to predict sound speed maps from input beamformed IQ data. The trained network estimates the sound speed and distribution of adipose, muscle, and liver tissue with a mean absolute error (MAE) of 23.6 m/s on in silico test data with thermal noise up to 5 dB. When tested on in vivo data, the predicted sound speeds appeared within expected values and showed clear tissue structures. These results show potential for using neural networks trained on full wave propagation physics for aberration correction of abdominal ultrasound scans, improving diagnosis and treatment of diseases like hepatocellular carcinoma.