Deep neural networks for ultrasound beamforming

Abstract

We trained deep neural networks (DNNs) to suppress off-axis scattering. The networks operated on sub-band ultrasound channel data in the frequency domain and a different network was trained for each frequency. The in-phase and quadrature components of the signals were treated as separate inputs to the network and the output structure was the same as the input. An inverse short-time Fourier transform (ISTFT) was used to reconstruct channel data. Training data was generated using simulations and the training data consisted of the responses from individual point targets. Performance was compared to standard delay-and-sum (DAS) beamforming using simulated a point target and anechoic cyst and a physical phantom anechoic cyst. For a simulated point target, the side lobes when using the DNN approach were about 80 dB below those of standard DAS. For a 5 mm diameter simulated anechoic cyst, the contrast ratio (CR), contrast-to-noise ratio (CNR), and speckle SNR (SNRs) were 50.9 dB, 4.7 dB, and 1.71, respectively, for the DNN approach and 32.0 dB, 4.8 dB, and 1.78, respectively, for standard DAS. For a 5 mm diameter anechoic cyst in a physical phantom, the CR, CNR, and SNRs were 47.5 dB, 4.76 dB, and 1.73, respectively, for the DNN approach and 26.7 dB, 5.5 dB, and 1.99, respectively, for standard DAS.