Research And Application On Ultrasound Imaging From Sub-sampled RF Data Using Deep Learning

Medical ultrasound imaging technology is a commonly used clinically assisted medical diagnosis method.The traditional B-mode ultrasound imaging equipment is bulky and expensive.With the continuous development of integrated circuits,portable B-mode ultrasonic diagnostic equipment can be implemented.In portable ultrasound imaging systems,there is an increasing demand for the reconstruction of high-quality images from a limited number of radio-frequency(RF)measurements due to receiver(Rx)or transmit(Xmit)event subsampling.However,due to the presence of side lobe artifacts from RF subsampling,the standard beamformer often produces blurry images with less contrast,which are unsuitable for diagnostic purposes.Therefore,trying to use deep learning methods to reconstruct high-quality ultrasound images from the limited radio-frequency data obtained by the sensor has important practical significance.Firstly,the research work in this paper simulates the scenario where portable devices reduce the number of Xmit events at different down-sampling rates.In the three imaging methods of focused line-scan imaging,far-focused pixel imaging and plane wave imaging,phantom and in vivo experiments are performed,the image quality quickly degrades as the number of measurement events decreases.To address this problem,here,we propose a deep-learning-based beamformer to generate significantly improved images over widely varying measurement conditions and Xmit events subsampling patterns.Secondly,Convolutional Neural Network and U-Net are used in our research as deep-learning-based data-driven deep beamformer.Research shows that the expressive ability of deep neural networks is mainly improved with the increase of network depth.Therefore,we only used components which are commonly apply in the field of deep learning to design our deep beamformer.The designed neural networks extract the main features of the RF data through the convolutional layer,use the pooling layer to compress the information,and the Re LU activation function layer provides efficient non-linear expression capabilities,and use the skip connection layer to avoid the problem of gradient disappearance caused by deepening the network.Finally,the proposed methods have outstanding performance in down-sampling RF data imaging task under all the three imaging modes.For focused line-scan imaging,the problem of blurred details in the down-sampling imaging result and the region of interest shifted due to down-sampling are successfully solved.For far-focused pixel imaging,the contrast of the image was enhanced.For plane wave imaging,the side lobes are suppressed and artifacts are eliminated.Comparing the reconstruction results of the two networks,U-Net has better stability and scene adaptability than convolutional neural networks.In summary,this article provides a deep learning imaging solution for downsampling RF data.The designed deep beamformer can effectively improve the quality of the ultrasound image reconstructed from down-sampling RF data.It can solve the problem of limited memory and insufficient computing power of portable devices.