Human skull profile and speed of sound estimation using pulse-echo ultrasound signals with deep learning

Abstract

Transcranial ultrasound brain imaging in human adults remains challenging, largely because human skulls cause severe phase aberration. Prior research has shown that phase aberration can be most accurately corrected if the skull profile (i.e., thickness distribution) and speed of sound are known apriori. We propose a deep-learning-based method to estimate the skull profile and sound speed from pulse-echo ultrasound signals. To demonstrate the feasibility of the method, k-wave simulations were performed to generate ultrasound radiofrequency (RF) signals backscattered from the skulls, using density and speed of sound maps derived from diagnostic CT scans of 5 ex vivo human skulls. An unfocused single-element transducer was simulated for ultrasound transmission and receiving at 0.75 MHz with 4-cycle tone bursts. A one-dimensional convolutional neural network (1D-CNN) model was designed and trained to predict the skull thickness and sound speed from RF signals. In an independent test set, the model yielded a mean absolute error of 0.3 mm for skull thickness prediction and 31 m/s for sound speed prediction. The result demonstrates that the deep learning method is capable of accurately estimating skull thickness and speed of sound, providing a potentially powerful tool for skull phase aberration correction in transcranial ultrasound brain imaging.