gCNR Regularization Improves Deep Neural Network Beamformers

Deep neural network (DNN) beamformers have gained traction in ultrasound because of their ability to approximate any non-linear function. In clutter suppression, DNN beamformers can be trained using paired synthetic data by minimizing a loss computed between the prediction and ground truth. In this work, we showed that minimizing a conventional Smooth-L<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</inf> loss does not correlate to improving the generalized contrast-to-noise (gCNR) ratio of the prediction. Instead, we propose to directly incorporate gCNR as a regularizer to the training loss and showed some improvement in synthetic data. Finally, we integrated gCNR regularization into an existing domain adaptation approach and achieved a gCNR gain of 0.0322 ± 0.0336 over delay-and-sum.