Improving Cardiac Ultrasound with a Semi-Supervised Deep Learning Beamformer

Deep learning beamformers have demonstrated the ability to remove a variety of artifacts from ultrasound images in recent years. Many of these algorithms operate on channel data, where in vivo ground truth data, free of any degradation, is unavailable. Much of the existing work estimates the ground truth distribution with synthetic data. Under this framework, the domain gap between the synthetic training data and in vivo test data limits the beamformer performance on inference. In this work, we introduce a multi-step, semi-supervised approach that leverages synthetic and in vivo data in training via cross-domain cycleGANs. We evaluate the intermediate generators with VCZ curves, and demonstrate that the beamformer trained with the proposed approach achieves a CNR gain of 2.72 ± 1.40 dB and gCNR gain of 0.284 ± 0.094 over delay-and-sum in a 32-frame test cine loop.