High Frame-Rate Ultrasound Imaging Using Deep Learning Beamforming

Unfocused transmit beams such as diverging waves (DW) and coherent compounding are essential in achieving higher volumetric frame rates in 3D ultrasound imaging. However, image quality loss that comes with the use of DW becomes an issue, especially when the number of transmits is small. We propose a deep learning beamforming method for eliminating some of the artifacts associated with DW imaging. We train a convolutional neural network to map the non-linear transformation between the aligned per-channel data from 11 DW transmits, before compounding, to the compounded per-channel data from 51 transmits. We include additional terms in our loss function such as the beamsum value and log-detected image pixel value to guide the learning in the desired direction. The neural network is trained and tested on simulation and in-vivo data. The final network successfully suppresses acoustic artifacts such as side lobe and clutter in the images obtained with 11 DW transmits.