Adversarial Training for Ultrasound Beamforming in Out-of-Distribution Scenarios

The out-of-distribution (OoD) generalization of ultrasound beamformers based on deep learning (DL) is typically addressed by increasing the variability in the training data. This often involves creating simulated datasets resembling real data or combining simulations with real data. However, incorporating data variations to enhance the robustness of beamformers is resource-intensive and non-systematic. In this work, adversarial training is explored as a more systematic approach to enhance the OoD generalization of a DL-based beamformer. Two training methods, standard and adversarial, were employed in a DL-based beamformer previously proposed in the literature. In both cases, the beamformer was trained with simulated data without attenuation and evaluated in OoD scenarios. The OoD scenarios involved a dataset simulated with an attenuation of 0.5 dB/cm-MHz and experimental data from an ATS Model 539 phantom. The results indicate that adversarial training improves the generalization of the standard-trained DL-based beamformer. The generalized contrast-to-noise Ratio of the inclusion improved from 0.59 +/- 0.13 with standard training to 0.72 +/- 0.10 with adversarial training in simulated attenuated data, and from 0.88 to 0.91 in physical phantom data. Moreover, artifacts produced by attenuation were significantly reduced by adversarial training in both scenarios. These findings highlight the potential of adversarial training as a method to systematically improve OoD robustness in DL-based ultrasound beamforming without the need for additional training scenarios.