基于DenseNet ECA的儿童异常胸片筛查
Screening Abnormal Chest X-Ray in Children Based on DenseNet ECA
段梦宇 1吴英飞 1袁贞明 1俞刚2
作者信息
- 1. 杭州师范大学信息科学与技术学院,浙江 杭州 311121;移动健康管理系统教育部工程研究中心,浙江 杭州 311121
- 2. 浙江大学医学院附属儿童医院数据信息部,浙江 杭州 310052;国家儿童健康与疾病临床医学研究中心,浙江 杭州 310052
- 折叠
摘要
胸片是筛查儿童肺部异常最常见、最容易获得的低成本成像方式.然而,在部分医疗资源匮乏地区,由于有经验的放射科医生数量稀少,导致胸片的解读效率低下,易造成对肺部异常患儿的漏诊、误诊.因此以儿童健康及异常胸片为研究对象,通过使用ECA注意力机制及PReLU激活函数对DenseNet进行改进,提出一种用于儿童异常胸片筛查任务的DenseNet ECA模型.实验结果表明,上述模型对于儿童健康、异常胸片的分类效果优于常用卷积神经网络模型,分类准确率、灵敏度、特异性分别可达 93.57%,91.47%,95.83%,参数量仅为 6.96M.上述模型能够帮助医生进行儿童异常胸片的预先筛查,可有效降低临床阅片压力,提高医生诊断效率.
Abstract
Chest X-ray is the most common and easily available low-cost imaging method for screening pulmonary abnormalities in children.However,in some areas where medical resources are scarce,due to the small number of experienced radiologists,the interpretation of chest X-ray is inefficient,which is easy to cause missed diag-nosis and misdiagnosis of children with pulmonary abnormalities.Therefore,this paper took children's health and ab-normal chest X-ray as the research object,improved DenseNet by using ECA attention mechanism and PReLU activa-tion function,and proposed a deep learning modelDenseNet ECA for children's abnormal chest X-ray screening task.The experimental results show that the classification effect of this model for children's healthy and abnormal chest X-rayis better than that of common convolutional neural networkmodels.The classification accuracy,sensitivity and spe-cificity can reach 93.57%,91.47%and 95.83%respectively,and the parameter quantity is only 6.96M.The model can help doctors to pre-screen abnormalchest X-ray of children,effectively reduce the pressure ofclinical film reading and improve the diagnostic efficiency of doctors.
关键词
儿童异常胸片筛查/注意力机制/仿真Key words
Abnormal chest x-ray screening in children/Attention mechanism/Simulation引用本文复制引用
基金项目
国家重点研发计划(.2019YFE0126200)
国家自然科学基金面上项目(62076218)
出版年
2024
NETL
NSTL
万方数据