3D Microbubble Localization with a Convolutional Neural Network for Super-Resolution Ultrasound Imaging

Expanding ultrasound localization microscopy from 2D to 3D data, a large amount of volume data has to be handled. Furthermore, for the reduction of measurement times it is of interest to be able to accurately localize microbubbles (MB) even for high MB concentrations. To meet these two requirements, we implemented and tested a convolutional neural network that is a U-net-oriented autoencoder for the 3D MB localization. It was trained and tested with Field II simulations where the settings were based on a Verasonics 8 MHz matrix array transducer. The trained network worked reliable up to concentrations of 1000 MB/cm
³
. The detection sensitivity of MB is higher than 90% and the false descovery rate at most 2%. With a mean absolute error of the MB localizations smaller than 0.3λ also a high accuracy was demonstrated.