Comparison of pulse sequences used for super-resolution ultrasound imaging with deep learning

Super-resolution ultrasound (SR-US) is a relatively recent development in the field of contrast-enhanced ultrasound (CEUS) imaging that has improved spatial resolution by up to an order of magnitude. Conventional processing algorithms can suffer from long computation times, preventing real-time data visualization and imaging as the SR-US maps are generated. Further, low signal-to-noise ratio (SNR) and contrast-to-tissue ratio (CTR) of CEUS images limits the number of microbubble (MB) contrast agents detected within a single ultrasound (US) image frame. This unnecessarily lengthens data acquisition times. Nonlinear contrast pulse sequences (CPS) have been shown to increase CTR and may allow more MBs to be detected in SR-US. The purpose of this study was to evaluate a deep learning approach with multiple nonlinear US pulse sequences for improving the MB detection step during SR-US image production. A three-dimensional (3D) convolutional neural network (3DCNN) model was used to perform spatiotemporal filtering (suppression of any residual tissue signal) and then MB detection in place of the more commonly used singular value decomposition (SVD) method. Three separate 3DCNN models were trained based on US images collected using different CPS pulse compositions, namely, one pulse (i.e. B-mode US), two pulse inversion (PI-2) and three pulse amplitude modulation with pulse inversion (AMPI-3). Validation of the 3DCNN models on simulated in silico data exceeded 99% accuracy for all CPS pulse compositions evaluated. Testing on in vivo data revealed that nonlinear CPS pulse compositions implemented on a programmable US scanner (Vantage 256, Verasonics Inc) detected up to 41% more MBs than when using B-mode US imaging and conventional SR-US processing. The average image processing frame rate was 51 Hz on a Nvidia GeForce 2080Ti GPU, which is promising for a realtime SR-US imaging modality.