Assessing and improving deep domain alignment in ultrasound via simulation diversity

Area of Science:

• Medical Imaging
• Acoustics
• Artificial Intelligence

Background:

• Ultrasound beamforming is essential for image formation.
• Conventional delay-and-sum (DAS) methods struggle with acoustic clutter.
• Deep neural network (DNN) beamformers trained on simulated data face domain gap challenges.

Purpose of the Study:

• To identify and quantify the contributions of reverberation and phase aberration to the domain mismatch in DNN beamforming.
• To evaluate methods for bridging the domain gap between simulated and in vivo ultrasound data.

Main Methods:

• Selective introduction of reverberation and phase aberration into simulated ultrasound data.
• Quantification of domain shift using KL divergence and Wasserstein-2 distance.
• Validation of domain gap reduction metrics with downstream beamformer performance.

Main Results:

• Including aberration and reverberation in simulations reduced the domain gap by 7.4% and 45%, respectively.
• A combined approach reduced the domain gap by 53%, while CycleGAN maps reduced it by 64%.
• CycleGAN performance was dependent on source distribution, with optimal results when both degradations were simulated.

Conclusions:

• Reverberation and phase aberration are significant contributors to the domain mismatch in DNN ultrasound beamforming.
• Simulating these artifacts improves the performance of DNN beamformers trained on synthetic data.
• Domain adaptation techniques like CycleGAN are effective but benefit from accurate simulation of image degradations.