Pan, Ying-Chun; Khan, Christopher M.; Berger, Matthew; Bryant, John M.; Lefevre, Ryan J.; Eagle, Susan S.; Byram, Brett C. (2026).听.听Ultrasonics, 167, 108193.听
Deep learning has shown promise for improving聽ultrasound beamforming聽(the process of combining ultrasound signals to create an image), but models trained on simulated data often perform poorly on real patient images because the two types of data differ. This study investigated the sources of that mismatch and how to reduce it. The authors focused on two common factors that degrade ultrasound images鈥攔everberation聽(unwanted echoes that create image artifacts) and聽phase aberration聽(distortions caused by sound traveling through different types of tissue). By selectively adding these effects to simulated data, they measured how closely the simulations matched real ultrasound images and assessed the impact on deep learning beamformer performance. Including reverberation reduced the gap between simulated and real data by 45%, while adding phase aberration reduced it by 7.4%. When both effects were included, the domain gap decreased by 53%, and applying a聽CycleGAN聽(an artificial intelligence method that translates images from one style to another) reduced it further to 64%. The study also found that the effectiveness of CycleGAN depended on the quality of the simulated training data, with the best beamforming performance achieved when both reverberation and phase aberration were included in the simulations. These findings suggest that creating more realistic simulated ultrasound data is an important step toward improving the performance of deep learning beamforming methods in real-world clinical settings.
Fig. 1.听Overview of DADNN method proposed by Tierney et al.听. A pair of CycleGAN (聽and聽) is trained between noisy simulation and noisy聽in vivo聽data. The forward map聽聽is then used to transform paired simulation data into the聽in vivo聽style and used as a target for the domain-adaptive,聽in vivo聽beamformer聽. To leverage paired simulation data,聽聽is jointly trained with a supervised beamformer聽聽with shared weights using the augmented feature mapping technique聽. More details are included in Section聽.