基于深度神经网络集成的肺炎图像检测方法研究  

Research on imaging detection method of pneumonia based on DNN integration

作　　者：杨洪兵 朱财林 毕卫云 冯桦 YANG Hongbing;ZHU Cailin;BI Weiyun;FENG Hua(Department of Radiology,the First Affiliated Hospital of the Air Force Medical University,Xi'an 710032,Shaanxi Province,China;Skills Training Center,the First Affiliated Hospital of the Air Force Medical University,Xi'an 710032,Shaanxi Province,China;Department of Thoracic Surgery,the First Affiliated Hospital of Xi'an Jiaotong University;Ultrasonic Diagnosis&Treatment Center,Xi'an International Medical Center Hospital)

机构地区：[1]空军军医大学第一附属医院放射科,西安710032 [2]空军军医大学第一附属医院临床技能培训中心,西安710032 [3]西安交通大学第一附属医院胸外科 [4]西安国际医学中心医院超声诊疗中心

出　　处：《中国数字医学》2023年第5期63-68,共6页China Digital Medicine

基　　金：国家自然科学基金(81402545)。

摘　　要：目的:基于深度神经网络模型捕捉肺炎特征,提高肺炎影像诊断的效率和准确率。方法:利用单个网络抽取符合单调性的类中心距离和最近邻标签分布特征,然后通过支持向量机(SVM)训练特征融合的集成模型,实现肺炎胸部X线片分类。结果:在公开数据集上,相比单一模型近94%的高准确率,集成方法提升近2%;在西京医院内部数据集上,集成方法实现5%左右的提升,准确率从75%左右提高至超过80%。结论:不同神经网络模型检测肺炎的绝对性能相近,且具有一定的互补性,基于SVM的集成学习可以有效提升肺炎影像诊断的准确率。Objective To improve the efficiency and accuracy of imaging diagnosis of pneumonia by capturing its features based on deep neural network(DNN)model.Methods A single network was used to extract the monotonic distribution features of class center distance and nearest-neighbor label,and then the integrated model was trained by support vector machine(SVM)to realize the classification of pneumonia X-ray films.Results On the open benchmark dataset,compared with the single DNN model(94%accuracy),the proposed integrated model improved the accuracy by 2%.On the internal dataset of Xijing Hospital,in which the single DNN model can only achieve around 75%accuracy,the proposed model realized an improvement of accuracy by around 5%to over 80%.Conclusion The absolute performance of different DNN models for pneumonia detection is similar,and they are complementary to each other to some extent.SVM-based integrated learning can effectively improve the accuracy of pneumonia detection.

关 键 词：肺炎检测 卷积神经网络 支持向量机 特征抽取 集成学习 

分 类 号：R816.4[医药卫生—放射医学] TP183[医药卫生—临床医学]