用于CT稀疏重建的对抗式多残差深度神经网络改进方法

Sparse-view CT reconstruction based on an adversarial multi-residual deep neural network

目的针对计算机断层成像(CT)稀疏重建过程中产生条状伪影的问题,本文提出了一种基于对抗式多残差深度神经网络的CT图像高精度稀疏重建方法。方法设计了一种耦合多残差和对抗机制的UNet网络,以含条状伪影图像和高精度图像作为训练样本,通过大规模训练数据,对该网络进行训练,使其具有压制条状伪影的能力。首先,利用滤波反投影(FBP)算法从稀疏投影中重建出含条状伪影的CT图像;其次,将其输入深度网络,通过网络压制条状伪影。最后,得到高精度的重建图像。结果实验结果表明,相比于总变差最小算法和现有的若干深度学习算法,提出新型网络重建出的图像精度更高,可以更好地压制条状伪影。结论将对抗机制和多残差机制引入经典的UNet网络,可以提高其压制CT图像条状伪影的能力。

Objective To solve the problem of severe streak artifacts in sparse-view computed tomography(CT)reconstruction,this paper proposes an adversarial multi-residual deep neural network to achieve high-quality sparse-view CT reconstruction.Methods A UNet that combines multi-residual and adversarial mechanism is designed.The network is trained through large-scale training data composed of streak artifact images and high-quality images to suppress streak artifacts.The filtered back projection(FBP)algorithm is used to reconstruct CT images with streak artifacts from sparse projections.These images are used as the input to the UNet,and streak artifacts are suppressed through this network to output high-quality images.Results The experimental results show that,compared with total variation minimization method and several existing deep learning algorithms,the image reconstructed by the proposed new network has higher accuracy and can suppress streak artifacts better.Conclusions Introducing the adversarial mechanism and the multi-residual mechanism into the classic UNet network can improve its ability to suppress streak artifacts in CT images.