基于深度学习的胸部X射线图像识别及分类模型研究  

Research on Chest X-ray Image Recognition and Classification Model based on Deep Learning

作　　者：张晓熙 王远涵 杨顷落 黄雪 余红梅 武淑琴 Zhang Xiaoxi;Wang Yuanhan;Yang Qingluo(Department of Health Statistics,School of Public Health,Shanxi Medical University(030000),Taiyuan)

机构地区：[1]山西医科大学公共卫生学院卫生统计学教研室,030000 [2]重大疾病风险评估山西省重点实验室 [3]山西医科大学基础医学院数学教研室

出　　处：《中国卫生统计》2024年第3期365-369,375,共6页Chinese Journal of Health Statistics

基　　金：国家自然科学基金面上项目(82273742)。

摘　　要：目的以深度学习中的卷积神经网络为基础搭建胸部X射线(chest X-ray,CXR)图像分类模型,为肺部疾病提供可靠的辅助诊断技术。方法经KAGGLE数据库收集新冠肺炎、轻度肺部感染、病毒性肺炎及正常的四种胸部X射线图片,按3∶1∶1的比例将数据随机划分成训练集,测试集和验证集;基于卷积神经网络架构搭建CXR图像分类模型,调节超参数对模型进行加强和优化;后通过混淆矩阵、准确率、灵敏度、K折交叉验证结果等指标对模型进行验证及评价。结果本研究模型对肺部影像图片的分类准确率为0.81、灵敏度为0.80、测试集和验证集损失值能够稳定在一个较低的水平。与相同迁移算法的模型相比,在测试数据集上的精确率、准确率、灵敏度、F1分值分别提高了1.7%、1.7%、1.3%、2.9%。结论此模型对于CXR图像的识别和分类的性能更强,可以更有效地应用于肺部疾病的辅助分析和判断。Objectives Building a chest X-ray(CXR)image classification model based on convolutional neural networks in deep learning,providing reliable auxiliary diagnostic techniques for lung diseases.Methods Four kinds of chest X-ray pictures of COVID-19,mild pulmonary infection,viral pneumonia and normal were collected through KAGGLE database,and the data were randomly divided into training set,test set and verification set according to 3∶1∶1 ratio.Building a CXR image classification model based on convolutional neural network architecture,adjusting hyperparameters to strengthen and optimize the model.Subsequently,the model was validated and evaluated using metrics such as confusion matrix,accuracy,sensitivity,and K-fold cross validation results.Results The classification accuracy of this research model for lung imaging images is 0.81,the sensitivity is 0.80,and the loss values of the test and validation sets can be stable at a relatively low level.Compared with models with the same migration algorithm,the accuracy,sensitivity,and F1 score on the test dataset were improved by 1.7%,1.7%,1.3%,and 2.9%,respectively.Conclusion This model has stronger recognition and classification performance for CXR images,and can be more effectively applied to auxiliary analysis and judgment of lung diseases.

关 键 词：肺部疾病 模型构建 深度学习 影像识别 

分 类 号：R195.1[医药卫生—卫生统计学]