Noise Suppression in Ultrasound Beamforming Using Convolutional Neural Networks

Abstract

Medical ultrasound is a noninvasive, affordable, portable, and real-time diagnostic modality that provides a cross-sectional view of tissues. Ultrasound beamforming is a widely used approach to acquire and process ultrasound data in a focused manner. However, noises from scatterers outside the main lobe of the beam degrade the beamformed images. Recently, frequency-domain multi-layer perceptrons (MLPs) prove effective in suppressing off-axis scattering and improving image contrast. This thesis extends the frequency-domain neural network approach to study the effectiveness of convolutional neural networks (CNNs). I propose a variety of convolutional architectures and investigate the contribution of the convolution operation. Preliminary results show that although hybrid convolutional- and full-connected neural networks can achieve a similar performance compared with MLPs, fully-convolutional neural networks do not perform well because they do not learn additional features. Instead, they approximate full connections by having a large effective receptive field.