Suppressing Off-axis Scattering Using Deep Neural Networks

摘要

We developed a method that uses deep neural networks (UNNs) to suppress off-axis scattering in ultrasound images. This approach operates in the frequency domain and networks were trained using the simulated responses from individual point targets. The network inputs consisted of the separated in-phase and quadrature components observed across the aperture of the array. The output had the same structure as the input and an inverse short-time Fourier transform was used to convert the processed data back to the time domain. In this work, we examined the noise handling characteristics of the DNN beamformer and also the relation between final image quality and the loss function for training networks.