Reconstruction for Diverging-Wave Imaging Using Deep Convolutional Neural Networks

Abstract

In recent years, diverging-wave (DW) ultrasound imaging has become a very promising methodology for cardiovascular imaging due to its high temporal resolution. However, if they are limited in number, DW transmits provide lower image quality compared with classical focused schemes. A conventional reconstruction approach consists in summing series of ultrasound signals coherently, at the expense of the frame rate. To deal with this limitation, we propose a convolutional neural networks (CNN) architecture for high-quality reconstruction of DW ultrasound images using a small number of transmissions. Given the spatially varying properties of DW images along depth, we adopted the inception model composed of the concatenation of multi-scale convolutional kernels. Incorporating inception modules aims at capturing different image features with multi-scale receptive fields. A mapping between low-quality images and corresponding high-quality compounded reconstruction was learned by training the network using in vitro and in vivo samples. The performance of the proposed approach was evaluated in terms of contrast-to-noise ratio and lateral resolution, and compared with standard compounding method and conventional CNN methods. The results demonstrate that our method could produce highquality images using only three DWs, yielding an image quality equivalent to the one obtained with standard compounding of 31 DWs and outperforming more conventional CNN architectures in terms of complexity, inference time and image quality.

Complex Convolutional Neural Networks for Ultrafast Ultrasound Imaging Reconstruction From In-Phase/Quadrature Signal

“…In particular, the second last layer is an inception layer, which is the concatenation of multi-scale convolution kernels. As demonstrated in [11], the inception layer used in conjunction with maxout activation allowed features from multiple receptive field sizes to be captured, which helped to address the specific geometry of DW.…”

“…CID-Net consists of the complex building components introduced in [10], which allowed incorporating complex numbers in general frameworks for training deep neural networks. Regarding the network architecture, CID-Net maintained the architecture of the inception for DW Network (ID-Net) [11], which has demonstrated the ability to reconstruct high-quality US images using RF data from DW acquisitions. We experimentally demonstrate that the CID-Net: i) yields the same image quality as that obtained from the RF-trained CNN; ii) outperformed the approach consisting in processing the real and imaginary parts of the I/Q signal separately.…”

Plane-Wave Ultrasound Beamforming: A Deep Learning Approach

“…Network (GAN) to achieve the quality of multi-focus ultrasound imaging using only a single focused transmission [19]. High quality image reconstruction of diverging-wave ultrasound imaging from a small number of transmissions based on CNNs is proposed in [20].…”

High-quality Low-dose CT Reconstruction Using Convolutional Neural Networks with Spatial and Channel Squeeze and Excitation

“…In recent years, deep learning has emerged as the de facto standard in various image processing problems. Inspired by the success of deep learning, many researchers have investigated deep learning methods for medical image reconstruction and achieved significant performance [4]- [9]. Lee et al [6] * indicates equal contribution proposed a convolution neural network (CNN) for the reconstruction of magnetic resonance (MR) images from accelerated MR acquisition.…”

“…Lee et al [6] * indicates equal contribution proposed a convolution neural network (CNN) for the reconstruction of magnetic resonance (MR) images from accelerated MR acquisition. In [9], Lu et al proposed a multi-scale CNN to improve diverging wave (DW) ultrasound (US) imaging, yielding high-quality US images while using a small number of DW transmissions. In CT literature, the systematic study of CNNs was first proposed in [10], where a CNN using directional wavelets was demonstrated to be more efficient in removing noises induced from low-dose CT. Jin et al [11] proposed to incorporate residual learning in the U-Net architecture, where large receptive fields were shown to be crucial to reduce globally distributed artifacts.…”