基于多尺度卷积神经网络的低剂量CT图像后处理  

Post-processing of low-dose CT images based on multi-scale convolutional neural network

作　　者：司菁菁[1,2] 张宁 赵熙[1] 程银波 SI Jingjing;ZHANG Ning;ZHAO Xi;CHENG Yinbo(School of Information Science and Engineering,Yanshan University,Qinhuangdao,Hebei 066004,China;Hebei Key Lab of Information Transmission and Signal Processing,Qinhuangdao,Hebei 066004,China;Ocean College,Hebei Agricultural University,Qinhuangdao,Hebei 066003,China)

机构地区：[1]燕山大学信息科学与工程学院,河北秦皇岛066004 [2]河北省信息传输与信号处理重点实验室,河北秦皇岛066004 [3]河北农业大学海洋学院,河北秦皇岛066003

出　　处：《燕山大学学报》2022年第4期362-370,共9页Journal of Yanshan University

基　　金：国家自然科学基金资助项目(60171429);河北省自然科学基金资助项目(F2021203027)。

摘　　要：低剂量CT通过降低X射线的剂量,减轻电离辐射对人体造成的伤害。然而,基于低剂量X射线重建出的CT图像通常含有大量的噪声与伪影,降低了图像的可读性。本文引入图像金字塔模型,结合残差结构,设计了一种基于多尺度双层卷积神经网络(Multi-scale Double-layer Convolutional Neural Network,MD-CNN)的低剂量CT图像后处理方案。该方案在图像金字塔模型的多尺度框架下,利用双层残差网络提取并融合不同尺度下的图像细节特征,提高重建图像的清晰度。利用实际人体胸部与腹部CT图像进行的实验表明,MD-CNN方案能够在有效去除低剂量CT图像中的噪声与伪影的同时,保留图像细节特征。与基于RED-CNN的低剂量CT图像后处理方案相比,MD-CNN方案重建图像的峰值信噪比平均提高了0.75 dB。Low-dose CT reduces the damage caused by ionizing radiation via reducing the dose of X rays.However,CT image reconstructed from low-dose X-ray usually contains a lot of noise and artifacts,which greatly reduce the readability of image.In this paper,a low-dose CT image post-processing scheme based on Multi-scale Double-layer Convolutional Neural Network(MD-CNN)is designed by combining the image pyramid model and residual structure.Under the multi-scale framework of the image pyramid model,detailed features of the image at different scales are extracted and fused by constructing a double-layer residual network,so as to improve the clarity of the reconstructed image.Experiments using real human chest and abdominal CT images show that MD-CNN can effectively remove the noise and artifacts in low-dose CT images,while retaining the detailed information.Compared with the low-dose CT image post-processing scheme based on RED-CNN,the peak signal-to-noise ratio of the reconstructed image of MD-CNN is improved by 0.75 dB on average.

关 键 词：低剂量CT 后处理 卷积神经网络 多尺度 

分 类 号：TN919.8[电子电信—通信与信息系统]