投审稿平台
投稿指南
下载专区
地  址:北京市海淀区中关村科学院
南路6号中国科学院计算所342号 [地图]
《计算机辅助设计与图形学学报》编辑部
邮政编码:100190
电  话:010-62562491
          010-62600342
订阅信息
ISSN      1003-9775
CN        11-2925/TP
邮发代号:82-456
单    价:80.00元
全年订价:960.00元
订阅电话:010-64017032
在线期刊

结合反卷积的CT图像超分辨重建网络

徐 军1,2), 刘 慧1,2)*, 郭 强1,2), 张彩明2,3,4)
1) (山东财经大学计算机科学与技术学院 济南 250014)2) (山东省数字媒体技术重点实验室 济南 250014)3) (山东大学计算机科学与技术学院 济南 250100)4) (山东高校未来智能计算协同创新中心 烟台 264025)
分类号: TP391.41 DOI: 10.3724/SP.J.1089.2018.17051
出版年,卷(期):页码: 2018 , 30 ( 11 ): 2084-2092 徐军
摘要: 医学图像的质量对于患者疾病的诊断、治疗乃至科学研究起着重要的作用. 然而, 受医疗设备和放射剂量等因素的影响, 医学CT图像的分辨率普遍较低. 为了实现医学CT图像超分辨重建, 提出一种结合反卷积的神经网络算法, 通过引入反卷积操作, 有效地建立了低/高分辨率图像之间端到端的映射. 首先选取肺部、脑部、心脏和脊椎等部位的1 500幅CT图像作为训练数据, 将训练数据下采样后输入网络模型; 然后建立正反卷积网络模型学习图像特征, 网络模型用caffe框架实现, 激活函数使用PReLU; 最后基于学习到的这些特征重建出高分辨率图像, 采用平均方法重建图像. 实验结果表明, 文中算法能够更好地重建出图像的轮廓和边缘纹理; 与已有算法相比, 所构建的4层网络结构在重建结果的峰值信噪比、结构相似性、信息熵及重建速度等性能指标上均取得了更好的效果.
关键词: CT图像; 超分辨重建; 卷积神经网络; 反卷积; PReLU
Super-Resolution Reconstruction of CT Images Using Neural Network Combined with Deconvolution
Xu Jun1,2), Liu Hui1,2)*, Guo Qiang1,2), and Zhang Caiming2,3, 4)
1) (School of Computer Science and Technology, Shandong University of Finance and Economics, Ji’nan 250014) 2) (Digital Media Technology Key Laboratory of Shandong Province, Ji’nan 250014)3) (School of Computer Science and Technology, Shandong University, Ji’nan 250100) 4) (Shandong Co-Innovation Center of Future Intelligent Computing, Yantai 264025)
abstract: The quality of medical images plays an important role in the diagnosis, treatment and even scientific research of patients’ diseases. However, due to the influence of medical equipment and radiation dose, the resolution of medical CT images is generally low. Therefore, this paper proposes a neural network algorithm combined with deconvolution for achieving super-resolution reconstruction of medical CT images. The proposed algorithm adds the operation of deconvolution, which effectively establishes the end-to-end mapping between low and high-resolution images. Firstly, 1 500 CT images of lung, brain, heart and spine are selected as training data, and they are down-sampled and then input into network model. Secondly, through the establishment of the convolution and deconvolution network model to learn image features, the network model is implemented using the caffe framework, and the activation function uses PReLU. Finally, the algorithm reconstructs high-resolution images using these features, where the image of reconstruction are averaged. The experimental results show that the proposed algorithm can reconstruct the contour and edge texture of the image better. Compared with the traditional super-resolution algorithm, the constructed four-layer network in this paper achieves better results in peak signal-to-noise ratio (PSNR), structural similarity (SSIM), information entropy (IE) and reconstruction speed.
keyword: CT image; super-resolution reconstruction; convolution neural network; deconvolution; PReLU
 
Copyright © 2004《计算机辅助设计与图形学学报》版权所有
电话:010-62600342 传真:010-62562491
E_mail:jcad@ict.ac.cn