Assessing and improving deep domain alignment in ultrasound via simulation diversity

Ultrasound beamforming plays a crucial role in the formation of an ultrasound image. The conventional delay-and-sum (DAS) method is efficient but susceptible to acoustic clutter which obscures imaging. While many clutter reduction techniques have emerged, recent efforts increasingly use deep neural network (DNN) beamformers trained on simulated data, which offers convenient access to ground truth. However, the domain gap between simulated training data and in vivo test data undermines model performance. Recent domain-adaptive work proposed to bridge the gap with a CycleGAN style-transfer map, but the limitation of the map and the source of domain mismatch remains poorly understood. This work seeks to identify the contribution of reverberation and phase aberration to the domain mismatch by selectively introducing them in simulation. We used KL divergence and Wasserstein-2 distance to quantify domain shift and validated their use with the downstream beamformer performance. Using the Wasserstein-2 distance between simulation data with no reverberation and phase aberration and in vivo data as reference, we showed that including aberration and reverberation in simulation reduced the domain gap by 7.4% and 45%, respectively. The domain gap was reduced by 53% when both sources of image degradation were included in simulation, and reduced by 64% with the use of CycleGAN maps. We further demonstrated that CycleGAN efficacy is dependent on source distribution and revealed that best beamformer results were achieved when reverberation and phase aberration were both present in simulation.