基于对抗式残差密集深度神经网络的CT稀疏重建

Sparse CT Reconstruction Based on Adversarial Residual Dense Deep Neural Network

针对计算机断层成像稀疏重建过程中产生条状伪影的问题,本文提出一种基于对抗式残差密集深度神经网络的CT图像高精度稀疏重建方法。设计一种耦合残差连接、密集连接、注意力机制和对抗机制的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,利用滤波反投影算法从稀疏投影中重建出含条状伪影的CT图像;接着,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。实验结果表明,相比于现有的若干深度学习算法,提出的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。

To solve the problem of severe streak artifacts in sparse-view computed tomography(CT)reconstruction,in this paper we propose a method which is based on the adversarial residual dense deep neural network to acquire high-quality sparse-view CT reconstruction.The UNet that combines residual connectioin,dense connection,adversarial mechanism and attention mechanism is designed,which is trained through largescale training data composed of streak artifact images and high-quality images to suppress streak artifacts.First,the filtered back projection(FBP) algorithm is used to reconstruct CT images with streak artifacts from sparse projections,then these images are inputed into the deep network,which can suppress streak artifacts to output high-quality images.The experimental results show that,compared with the existing deep learning algorithms,the image reconstructed by the proposed new network possesses higher accuracy and can suppress streak artifacts better.