Deep Learning Based Ultrasound Tomography for Real-Time Brain Imaging

Ultrasound Computed Tomography (USCT) is an innovative technique that enhances the accuracy of traditional ultrasound. However, conventional USCT reconstruction methods typically depend on iterative algorithms to determine the optimal sound speed distribution that matches ultrasound signals. These algorithms are unsuitable for real-time reconstructions owing to their iterative nature. Consequently, despite offering a higher resolution, USCT lacks a crucial feature compared to traditional ultrasound. To enhance the availability of real-time imaging at USCT, we propose a neural network as an end-to-end solution for generating segmented brain tissue maps directly from the recorded sensor data. Our Convolutional Neural Network (CNN) employs 1D convolutions to efficiently process transmitter signals, enabling fast and accurate predictions of segmented tissue maps. The neural network was trained and tested using simulation data produced by the open-source acoustic wave solver, K-wave. The phantoms for the forward simulation were randomly generated to mimic horizontal sections of the human brain. The proposed model demonstrated high accuracy in generating segmented tissue maps while significantly reducing the reconstruction process time to less than one second.