摘要
目的:基于深度学习根据儿童胸部X光正位数字影像构建肺炎自动判别模型,辅助临床诊断,提高影像诊断效率。方法:首先通过选取公开数据集5856张儿童胸片(肺炎4273张,正常1583张),分为训练集、验证集和测试集,基于Resnet-50神经网络构建儿童肺炎自动判别模型,利用验证集选取最优模型,在测试集上做内部独立验证。进一步收集6家医疗单位共611张儿童胸片(肺炎300张,正常311张)进行外部验证,并根据验证结果对模型进行微调后再次测试,使模型更适合临床使用。结果:基于深度学习技术和公开数据集数据构建儿童肺炎自动判别模型,准确率为98.48%,精确率为99.54%,召回率为98.81%,F1-score为98.86%,AUC为0.999。外部验证初始结果准确率为59.90%,选用部分外部验证数据微调模型后,独立测试准确度提升至85.00%。结论:基于深度学习根据公开数据集构建肺炎自动判别模型具有可行性,准确率达98.48%,在实际临床使用时应根据具体使用条件选取适量数据集对模型进行微调。
Objective To construct a deep learning-based model for automatically detecting pneumonia according to the digital ortho-images of children's chest X-ray for assisting clinical diagnosis and improving the efficiency of image diagnosis.Methods A total of 5856 pediatric chest radiographs,including 4273 chest radiographs of pneumonia and 1583 normal chest radiographs,were selected from the public data set and divided into training set,verification set and test set.A model for the automated pediatric pneumonia detection was constructed based on Resnet-50.The validation set was used for selecting the optimal model,and the test set for carrying out internal independent validation.In addition,611 pediatric chest radiographs,including 300 chest radiographs of pneumonia and 311 normal chest radiographs,were further collected from 6 medical units for external validation,and the model was fine-tuned according to validation results and then tested again to make it more suitable for clinical application.Results An automated detection model for pediatric pneumonia was successfully constructed using deep learning technology and public data set.The accuracy,precision,recall,F1-score and AUC of the model were 98.48%,99.54%,98.81%,98.86%and 0.999,respectively.After fine-tuning the model with some external validation data,the accuracy of the independent test was improved from 59.90%(preliminary external validation)to 85.00%(independent test).Conclusion It is feasible to construct an automated pneumonia detection model using deep learning and public data set,and the accuracy of the model can reach 98.48%.In practice,the model should be fine-tuned by selecting the appropriate data set according to the specific conditions.
作者
董芳芬
陈群
李诺兮
徐本华
李小波
DONG Fangfen;CHEN Qun;LI Nuoxi;XU Benhua;LI Xiaobo(Clinical Research Center for Radiology and Radiotherapy for Digestive,Hematological and Breast Malignancies of Fujian Province,Fujian Key Laboratory of Intelligent Imaging and Precision Radiotherapy for Tumors,Department of Radiation Oncology,Fujian Medical University Union Hospital,Fuzhou 350001,China;School of Medical Imaging,Fujian Medical University,Fuzhou 350004,China;School of Computer Science,Northwestern Polytechnical University,Xi'an 710072,China;Department of Engineering Physics,Tsinghua University,Beijing 100084,China)
出处
《中国医学物理学杂志》
CSCD
2022年第12期1579-1584,共6页
Chinese Journal of Medical Physics
基金
福建省科技厅(高校产学研)项目(2020Y4010)。
关键词
儿童
肺炎
深度学习
神经网络
children
pneumonia
deep learning
neural network
