基于ResNet深度模型的SPECT肺灌注图像分类  

Classifying SPECT Lung Perfusion Images Based on ResNet Models

作　　者：增思涛 曹永春[2] 林强 满正行[2] 邓涛[2] 王茸 ZENG Si-tao;CAO Yong-chun;LIN Qiang;MAN Zheng-xing;DENG Tao;WANG Rong(Key Laboratory of Chinese Ethnic Languages and Information Technology of Ministry of Education,Northwest Minzu University,Lanzhou 730030,China;Key Laboratory of Dynamic Streaming Data Computing and Application,Northwest Minzu University,Lanzhou 730030,China;Department of Nuclear Medicine,Gansu Provincial Hospital,Lanzhou 730030,China)

机构地区：[1]西北民族大学中国民族语言文字信息技术教育部重点实验室,甘肃兰州730030 [2]西北民族大学动态流数据计算与应用实验室,甘肃兰州730030 [3]甘肃省人民医院核医学科,甘肃兰州730020

出　　处：《西北民族大学学报（自然科学版）》2021年第2期27-35,共9页Journal of Northwest Minzu University(Natural Science)

基　　金：国家自然科学基金项目(61562075);西北民族大学中央高校基本科研业务费专项资金资助研究生项目(Yxm2019115)。

摘　　要：SPECT肺灌注是重要的功能成像技术,它能够以非入侵方式捕获肺部的功能病变,已经成为肺栓塞等疾病的重要临床检测手段.为有效支撑肺栓塞疾病的自动诊断,研究并构建基于ResNet深度网络的SPECT肺灌注图像分类器.首先采用归一化技术将原始SPECT肺灌注文件转换为SPECT图像,并应用图像平移和旋转等技术对图像进行预处理,以扩展SPECT图像样本数量,然后在标准ResNet-50深度模型的基础上引入特征融合和迁移学习技术,构建可靠的SPECT肺灌注图像分类器,最后使用一组真实的SPECT肺灌注图像对构建的分类器进行测试,实验结果表明构建的分类器可有效检测肺栓塞病变,获得的分类准确率超过95.5%.SPECT pulmonary perfusion imaging is a key functional imaging technology,which has the potential to capture pulmonary diseases in a non-invasive manner.SPECT pulmonary perfusion imaging has been become the important clinical detection method for obstructive pulmonary disease and other relative diseases.To effectively support the automatic diagnosis of obstructive pulmonary disease,a ResNet based deep classifier is developed for classifying SPECT pulmonary perfusion images.First,normalization technique is used to transform the original data files of SPECT pulmonary perfusion imaging,followed by image preprocessing operations including translation and rotation that enable to augment the original dataset.Second,feature fusion and transfer learning techniques are introduced to the standard ResNet-50 model to develop deep classifier that is able to answer that whether or not an obstructive pulmonary disease presents in the given SPECT pulmonary perfusion image.Last,experimental evaluation conducted on a group of real-world SPECT pulmonary perfusion images has shown that our classifier is effective and workable for identifying obstructive pulmonary diseases in SPECT pulmonary perfusion images,obtaining a value 95.5%for accuracy evaluation metric.

关 键 词：图像分类 核素肺灌注成像 深度学习 残差网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]