Image Quality-Based Regularization for Deep Network Ultrasound Beamforming

Abstract

Deep neural networks (DNNs) have previously been used to perform adaptive beamforming and improve image quality compared to conventional delay-and-sum (DAS). Although effective, low training validation loss is often not correlated to improved image quality, making model selection difficult. This discrepancy is due to these DNNs being optimized to perform an intermediate beamforming step instead of being optimized to enhance image quality on fully reconstructed images. Therefore, selecting model hyperparameters that produce optimal image quality has needed to be random and exhaustive. To address this problem, we propose a beamforming-relevant, end-to-end training scheme by using contrast-to-noise ratio (CNR) as a form of regularization. We compare a CNR-regularized DNN to a conventional DNN as well as DAS. When tested on simulated anechoic cysts, CNR-regularization resulted in 46% and 33% increases in CNR compared to the conventional DNN and DAS, respectively. When tested on in vivo data, CNR-regularization resulted in 68% and 25% increases in CNR compared to conventional DNN and DAS, respectively.