基于深度学习的胸部平片人工智能自动诊断模型的设计及研究  

The design and research of automatic diagnosis model of chest X-rays film based on deep learning

作　　者：辛小燕[1] 张帅[2] 蒋宛平 汤子洋 佟琪[1] 张艳秋 李丹燕[1] 张建[2] 张冰[1] XIN Xiaoyan;ZHANG Shuai;JIANG Wanping;TANG Ziyang;TONG Qi;ZHANG Yanqiu;LI Danyan;ZHANG Jian;ZHANG Bing(Department of Radiology,Nanjing Drum Tower Hospital,Nanjing University Medical School,Nanjing 210008,China;School of Physics,Collaborative Innovation Center of Advanced Microstructures,Nanjing University,Nanjing 210008,China;Nanjing Drum Tower Hospital,the Affiliated Hospital of Nanjing Medical University,Nanjing 210008,China)

机构地区：[1]南京大学医学院附属鼓楼医院医学影像科,江苏南京210008 [2]南京大学人工微结构协同创新中心,江苏南京210008 [3]南京医科大学鼓楼临床医学院,江苏南京210008

出　　处：《实用放射学杂志》2020年第5期822-825,共4页Journal of Practical Radiology

基　　金：江苏省“六大人才高峰”高层次人才项目(2016-WSN-160);南京市卫生科技发展专项基金项目医药卫生科研课题(YKK18062).

摘　　要：目的基于深度学习的方法,建立胸部X线平片人工智能诊断生成系统,并对其诊断性能进行初步研究.方法采用人工智能深度学习的方法,建立胸部X线平片人工智能自动诊断生成系统,预先用已经发布的数据集ChestX-ray8在监督模式下训练了编码器.然后用测试集NJDTH8进行测试,此测试集包括了12219份图像和同样数量的中文报告,基于一致性的图形描述(CIDEr)评分法对训练集和测试集数据进行评估.最后随机抽取100份由影像科医生给出诊断报告数据和100份有由模型自动生成诊断报告数据,随机排序后,再由2位高年资的影像科医生来判读评分.结果在训练集和测试集中,CIDEr评分法结果显示训练损失持续下降,验证损失在第10轮训练中达到饱和值5.8.在随机的200份报告中,模型生成报告的5分率达到72%,基本达到中等年资水平的影像诊断医生水平77%.结论本研究基于深度卷积神经网络和结合注意力机制的递归神经网络开发的胸部X线平片人工智能自动诊断模型,能够对胸部X线片自动生成影像学诊断报告,且诊断性能较好,可以继续进一步使用和深入研究.Objective To establish an artificial itelligence diagnostic generation system for chest X-rays(CXR)film,which was based on deep learning,and study on its diagnostic efficiency preliminarily.Methods First,this study developed an artificial itelligent automatic diagnostic system which can automatically generate radiological reports for given CXR scans with deep learning approach.The encoder has been trained under ChestX-ray8 dataset with the mode of supervision,it also has been tested by the testing dataset which named NJDTH8 including 12219 images and the same number of Chinese reports.And the quality of the generated reports from training dataset and testing dataset was estimated with CIDEr scores.Finally,100 diagnostic reports given by radiologists and 100 diagnostic reports generated automatically by the model were randomly selected.After randomly ordering,two senior radiologists were asked to judge the scores.Results In the training set and test set,the results of CIDEr score showed that the training loss continues to decline,and the verification loss reaches the saturation value of 5.8 in the tenth round of training.In 200 random reports,the 5-point rate of model generation report reached 72%,bsically reaching 77%of the level of medium-aged radiologistsConclusion The artificial intelligence automatic diagnosis generation model for plain CXR based on deep learning has a good diagnostic function,which worth further use and in-depth study.

关 键 词：胸部X线检查 机器学习 深度学习 神经网络 注意机制 

分 类 号：R816.4[医药卫生—放射医学] R814.49[医药卫生—临床医学]