Evaluating the Robustness of Ultrasound Beamforming with Deep Neural Networks

Abstract

We evaluated the robustness of deep neural network (DNN) beamforming to noise, gross sound speed errors, and phase aberration. Training data was generated using simulations and the training data consisted of the responses from point target responses. Performance was compared to standard delay-and-sum (DAS). When the channel SNR was 10 dB, the CNR for DNN and DAS beamforming were 5.4±0.1 dB and 5.0±0.1 dB, respectively. When the channel SNR was −10 dB, the CNR for DNN and DAS beamforming were 4.1±0.3 dB and 1.9±0.1 dB, respectively. When the assumed sound speed was 10% larger than the actual sound speed, the CNR for DNN and DAS beamforming were 4.9±0.2 dB and 4.7±0.3 dB, respectively. When a near field phase screen aberration profile with FWHM of 2.5 mm and RMS of 30 ns was introduced, the CNR for DNN and DAS beamforming were 2.9±0.5 dB and 0.8±1.1 dB, respectively. Overall, these results show that DNN beamforming was more robust to the examined sources of image degradation than DAS.