Locating the Interface between Different Media Based on Matrix Ultrasonic Sensor Data Using Convolutional Neural Networks

Introduction. The study focuses on modelling the process of ultrasound medical examination in a heterogeneous environment with regions of significantly different sound speeds. Such scenarios typically arise when visualizing brain structures through the skull. The aim of this work is to compare possible approaches to determining the interface between acoustically contrasting media using convolutional neural networks.Materials and Methods. Numerical modelling of the direct problem is performed, obtaining synthetic calculated ultrasonic images based on known geometry and rheology of the area as well as sensor parameters. The calculated images reproduce distortions and artifacts typical for setups involving the skull wall. Convolutional neural networks of 2D and 3D structures following the UNet architecture are used to solve the inverse problem of determining the interface between media based on a sensor signal. The networks are trained on computational datasets and then tested on individual samples not used in training.Results. Numerical B-scans for characteristic setups were obtained. The possibility of localizing the aberrator boundary with good quality for both 2D and 3D convolutional networks was demonstrated. A higher quality result was obtained for the 3D network in the presence of significant noise and artifacts in the input data. It was established that the 3D architecture network can provide the shape of the interface between media in 0.1 seconds.Discussion and Conclusions. The results can be used for the development of transcranial ultrasound technologies. Rapid localization of the skull boundary can be incorporated into imaging algorithms to compensate for distortions caused by differences in sound velocities in bone and soft tissues.