基于深度卷积神经网络的异常肺部CT图像初筛平台的设计与临床实践  

Design and clinical practice of abnormal lung CT image screening platform based on deep convolutional neural network

作　　者：刘红光 张行坤 刘诚 李伟凯 LIU Hongguang;ZHANG Xingkun;LIU Cheng;LI Weikai(Qingdao Central Hospital Affiliated to Qingdao University,Qingdao 266042,China;Department of Medical Imaging,Qingdao Women's and Children's Hospital Affiliated to Qingdao University,Qingdao 266034,China)

机构地区：[1]青岛大学附属青岛市中心医院,山东青岛266042 [2]青岛大学附属青岛市妇女儿童医院医学影像科,山东青岛266034

出　　处：《电子设计工程》2023年第23期184-188,195,共6页Electronic Design Engineering

基　　金：青岛市2021年度医药卫生科研计划项目(2021-WJZD075)。

摘　　要：针对肺部疾病初筛实践中CT图像异常先验判读机制的效能性问题,设计了一种基于深度卷积神经网络的异常肺部CT图像初筛平台并开展了临床实践验证。该平台使用CT扫描设备采集肺部CT图像,形成数据池;利用深度卷积神经网络对数据池训练集进行肺部异常特征学习,构建时间正序下的肺部CT图像异常特征全息感知机制;在Inception V3模型中结合Google Net网络结构,实现参数自动调整优化,达到对肺部疾病精准初筛的目的。使用青岛市中心医院临床肺部CT图像集进行实验,结果表明,该平台在肺部CT图像异常判读均值准确率达96.52%,肺部疾病初筛均值有效率达93.07%,可实现肺部疾病的初筛工作。Aiming at the efficiency of the abnormal CT image a priori interpretation mechanism in the practice of preliminary screening of lung diseases,a platform for preliminary screening of abnormal lung CT images based on deep convolutional neural network was designed and verified in clinical practice.The platform uses CT scanning equipment to collect lung CT images to form a data pool;uses a deep convolutional neural network to learn lung abnormality features from the data pool training set,and builds a holographic perception mechanism for abnormal features of lung CT images in positive time sequence;The Google Net network structure is introduced into the Inception V3 model,and the parameters are automatically adjusted and optimized during the training process,so as to achieve the purpose of accurate primary screening of lung diseases.Experiments were carried out using the clinical lung CT image set of Qingdao Central Hospital.The results showed that the average accuracy rate of the platform in the abnormal interpretation of lung CT images was 96.52%,and the average effective rate of primary screening of lung diseases was 93.07%,which can achieve lung diseases preliminary screening work.

关 键 词：肺部CT图像 异常自主判读 深度卷积网络 Google Net模型 疾病初筛 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]