Pan, Ying-Chun; Khan, Christopher; Lefevre, Ryan; Eagle, Susan; Byram, Brett. “GCNR Regularization Improves Deep Neural Network Beamformers.” IEEE Ultrasonics, Ferroelectrics, and Frequency Control Joint Symposium, UFFC-JS 2024 – Proceedings, 2024, https://doi.org/10.1109/UFFC-JS60046.2024.10793808.
Deep neural network (DNN) beamformers are becoming increasingly popular in ultrasound imaging due to their ability to handle complex, non-linear functions. In the context of clutter suppression, these DNN beamformers are typically trained using synthetic data, where the model learns by comparing its predictions to a known ground truth.
In this study, we found that the commonly used Smooth-L1 loss function does not necessarily lead to improvements in the generalized contrast-to-noise ratio (gCNR), a key measure of image quality. To address this, we introduced a new training approach that directly incorporates gCNR as a regularizer in the loss function. Our results showed that this method improved performance on synthetic data. Furthermore, when we integrated gCNR regularization into an existing domain adaptation technique, we achieved a measurable improvement in gCNR, with a gain of 0.0322 ± 0.0336 compared to the traditional delay-and-sum method.
