WO2022000360A1 - 一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法 - Google Patents

一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法 Download PDF

Info

Publication number
WO2022000360A1
WO2022000360A1 PCT/CN2020/099674 CN2020099674W WO2022000360A1 WO 2022000360 A1 WO2022000360 A1 WO 2022000360A1 CN 2020099674 W CN2020099674 W CN 2020099674W WO 2022000360 A1 WO2022000360 A1 WO 2022000360A1
Authority
WO
WIPO (PCT)
Prior art keywords
chest
pneumoconiosis
ray
diagnosis
radiographs
Prior art date
Application number
PCT/CN2020/099674
Other languages
English (en)
French (fr)
Inventor
李智民
张镏琢
张乃兴
罗军
杨新跃
纪祥
张�雄
Original Assignee
深圳市职业病防治院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳市职业病防治院 filed Critical 深圳市职业病防治院
Priority to PCT/CN2020/099674 priority Critical patent/WO2022000360A1/zh
Publication of WO2022000360A1 publication Critical patent/WO2022000360A1/zh

Links

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H15/00—ICT specially adapted for medical reports, e.g. generation or transmission thereof
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/20—ICT specially adapted for the handling or processing of medical images for handling medical images, e.g. DICOM, HL7 or PACS
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/50—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for simulation or modelling of medical disorders

Definitions

  • the invention relates to a method for occupational health examination screening or auxiliary diagnosis of pneumoconiosis, in particular to technologies such as statistical efficacy estimation and selection of sample numbers, parallel and serial combination of pneumoconiosis chest radiograph reading and labeling, and artificial intelligence model verification.
  • Pneumoconiosis is a lung disease mainly caused by diffuse fibrosis of lung tissue caused by long-term inhalation of productive mineral dust and retention in the lungs during occupational activities. It is the most serious and common occupational disease in the world.
  • pneumoconiosis is mainly based on X-ray high kilovolt or digital photography (DR) anterior chest X-ray as the basis for screening or diagnosis.
  • DR digital photography
  • the purpose of the present invention is to provide a method for selecting samples by using statistical power estimation method and combining parallel and series reading and labeling, to establish an artificial intelligence screening system for pneumoconiosis, and to obtain ideal results through verification. It enables artificial intelligence to improve the accuracy of screening and diagnosis, as well as work efficiency, in the occupational health examination and screening of dust-exposed workers and auxiliary diagnosis of pneumoconiosis.
  • a method for artificial intelligence screening of pneumoconiosis chest X-ray selection, labeling and verification includes a computer hardware system and a control software system, and the computer hardware system includes a central processing unit, Memory, input and output control system; the method includes the following steps:
  • DR digital X-ray
  • Shape and size of small shadows record the shape and size of small shadows, mark them with p, q, r, s, t, u or a combination, such as p/p, p/q, q/q, q/p , p/s, s/s, s/t, t/t, t/s, etc.;
  • Distribution range The above features in the chest X-ray are marked in the corresponding lung areas of the two lung fields to indicate the distribution range;
  • the method of combining parallel and series that is, double-blind and collective reading method, is used to read and label the films, so as to improve the accuracy of standard films;
  • group A and group B should jointly read the films, or choose an artificial person with the same conditions, that is, group C to read the films, and those with the same diagnosis of pneumoconiosis and the same staging results are regarded as standard films, and it is determined that the film does not exceed 2/3 consistent chest X-rays were excluded, and the interval between each reading was more than 2 weeks;
  • the standard radiographs are respectively stored in the computer normal chest radiograph standard database, the first-stage chest radiograph standard database, the second-stage chest radiograph standard database, and the third-stage chest radiograph standard database;
  • Convolutional neural network method one of the deep learning methods, is used to learn and train DR chest X-ray DICOM images, including normal chest X-ray standard database, first-stage chest X-ray standard database, second-stage chest X-ray standard database, and third-stage chest X-ray standard database
  • DR chest X-ray DICOM images including normal chest X-ray standard database, first-stage chest X-ray standard database, second-stage chest X-ray standard database, and third-stage chest X-ray standard database
  • the size of the chest radiograph was first adjusted to be smaller, and the two lung fields of the chest radiograph were cut out and divided into 6 areas to improve the resolution of the chest radiograph, and the chest radiograph was readjusted to the original size. , which is convenient for the computer to recognize the image;
  • the model consists of three components: input layer, hidden layer and output layer; among them, the input layer is used to accept the input of DR chest X-ray DICOM image, and pass it to the hidden layer, and finally pass the signal to the output layer; finally, pneumoconiosis is obtained. and staging forecast results;
  • the computer enters the chest radiograph for verification, calculates and identifies the characteristics of the chest radiograph, the shape and size of the small shadow, the density, the overall density, the distribution range, the large shadow, and the small shadow aggregation, and stores them in the chest to be checked. slice database;
  • the computer inputs the characteristics of the chest radiograph in the database of chest radiographs to be checked into the primary model for screening pneumoconiosis, obtains the results after calculation, and generates a diagnosis report of pneumoconiosis and staging;
  • the indicators are: true negative (TN), false negative (FN), false positive (FP), true positive (TP);
  • the accuracy of screening for pneumoconiosis is 97% or more
  • the accuracy of staging is 80% or more
  • the sensitivity is 0.98
  • the specificity is 0.97.
  • the invention can accurately and efficiently mark pneumoconiosis chest radiographs, and is used for the establishment of artificial intelligence screening pneumoconiosis models, so that artificial intelligence can improve screening and diagnosis in the occupational health examination screening and auxiliary diagnosis of pneumoconiosis of workers exposed to dust. Accuracy and improve work efficiency.
  • a chest radiograph selection, labeling and verification method for artificial intelligence screening for pneumoconiosis includes a computer hardware system and a control software system
  • the computer hardware system includes a central processing unit, a memory, and an input and output control system; it is characterized in that the The method includes the following steps:
  • the selection of chest radiographs in the present invention is based on the statistical power estimation method to determine the number of selected chest radiographs.
  • the ROC curve, the receiver operating characteristic curve, is a coordinate-schematic analysis tool used to select the best classification model and discard the second best classification model.
  • AUC Absolute under the Curve of ROC
  • the calculated statistical power was 0.89 for N 1 normal chest X-ray comparison samples and N 2 pneumoconiosis stage I chest X-ray samples. If the true AUC is 0.90, then the model is more accurate enough to distinguish normal chest X-ray samples from pneumoconiosis first-stage chest X-ray samples. The AUC is significantly higher than 0.85 (significance level is 0.05).
  • the calculated statistical power was 0.90 for the N 2 chest radiograph samples of stage I pneumoconiosis and the N 3 chest radiograph samples of stage II pneumoconiosis. If the true AUC is 0.88, then the model is more accurate enough to distinguish normal chest X-ray samples from pneumoconiosis first-stage chest X-ray samples. The AUC is significantly higher than 0.85 (significance level is 0.05).
  • Small shadow shape and size record the shape and size of small shadows, mark them with p, q, r, s, t, u or a combination, such as p/p, p/q, q/q, q/p , p/s, s/s, s/t, t/t, t/s, etc.;
  • Intensity graded and marked according to three grades and twelve small grades, expressed as follows: 0/-, 0/0, 0/1 is grade 0; 1/0, 1/1, 1/2 is grade 1; 2/1, 2/2, 2/3 are grade 2; 3/2, 3/3, 3/+ are grade 3;
  • Small shadow clusters denoted by 9, and marked with the lung area, and marked with green line drawing on the chest X-ray.
  • Distribution range The above features in the chest X-ray were marked in the corresponding lung areas of the two lung fields to indicate the distribution range.
  • group A and group B For the chest X-rays with inconsistent reading results between the two groups, group A and group B read the films together, or select an artificial person with the same conditions, that is, group C to read the films.
  • the diagnosis of pneumoconiosis and the staging results are the same as the standard films, and it is determined that the film does not exceed 2/3 consistent chest X-rays were excluded, and the interval between each reading was more than 2 weeks;
  • the standard radiographs were stored in the computer normal chest radiograph standard database, the first-stage chest radiograph standard database, the second-stage chest radiograph standard database, and the third-phase chest radiograph standard database.
  • Convolutional neural network one of the deep learning methods, is used to label the features of the chest radiographs in the normal chest radiograph standard database, the first-stage chest radiograph standard database, the second-stage chest radiograph standard database, and the third-stage chest radiograph standard database, including small Shadow shape and size, density, overall density, distribution range, large shadow, small shadow aggregation, etc., to learn and train, and to establish an artificial intelligence screening pneumoconiosis model.
  • the model consists of three components: input layer, hidden layer and output layer; wherein, the input layer is used to accept the input of the original DICOM image and pass it to the hidden layer.
  • the hidden layer can be set up with neurons of different layers, or adopt different arrangements, and finally transmit the signal to the output layer.
  • the output layer summarizes the signals of the last layer of the hidden layer to obtain the prediction result of the pneumoconiosis chest X-ray staging.
  • the computer enters the chest radiograph for verification and stores it in the database of the chest radiograph to be checked, and the computer identifies the characteristics of the chest radiograph, the shape and size of the small shadow, the density, the overall density, the distribution range, the large shadow, and the small shadow aggregation. , and respectively stored in the database of chest X-ray image features to be checked;
  • the computer inputs the characteristics of the chest X-ray in the image feature database of the chest X-ray to be checked into the initial model for screening pneumoconiosis, and after calculation, gives the diagnosis result, and generates a diagnosis report with the above-mentioned data;
  • the computer stores the unchecked chest X-ray and the diagnosis result in the diagnosis and treatment database.
  • the indicators are: true negative (TN), false negative (FN), false positive (FP), true positive (TP);
  • the screening accuracy rate is 97% or more
  • the staging accuracy rate is 80% or more
  • the sensitivity is 0.98
  • the specificity is 0.97.

Abstract

一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法；包括计算机硬件系统及控制软件系统，计算机硬件系统包括中央处理器、存储器、输入输出控制系统；其方法包括以下步骤：基础数据库的建立、计算与建模、对人工智能模型验证。该方法可准确和高效的对尘肺病胸片进行标注，并用于人工智能筛查尘肺模型的建立，使人工智能在接触粉尘工人职业健康检查筛查和尘肺病辅助诊断中，提高筛查和诊断准确率、提高工作效率。

Description

一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法 技术领域
本发明涉及职业健康检查筛查或辅助诊断尘肺病方法，特别涉及统计功效估计遴选样本数、并联与串联相结合的尘肺病胸片读片标注、人工智能模型验证等技术。
背景技术
尘肺病是在职业活动中长期吸入生产性矿物性粉尘并在肺内潴留而引起的以肺组织弥漫性纤维化为主的肺部疾病，是目前世界上危害最严重和最常见的职业病。
目前尘肺病主要以X射线高千伏或数字化摄影(DR)后前位胸片作为筛查或诊断依据。长期的实践和研究证明，在尘肺病体检或诊断中，读片差异是客观存在的。在遵循相同的诊断原则及读片条件下，不同读片者或同一读片者前后判断的结果均不尽相同，这种差异严重影响尘肺病筛查或诊断的准确性，极易产生误诊或漏诊。通过检索1990年1月1日至2018年8月10日中国关于尘肺病诊断差异的文献资料并进行系统分析，结果显示，对尘肺病诊断的总体符合率在15.0％至64.0％之间。
近年来，人工智能在医疗卫生领域的研究与应用发展迅速。人工智能以强大的学习能力、计算能力、识别能力作为诊疗辅助手段发挥了重 要作用，尤其是利用影像学数据对肺部疾病进行诊断发展迅速。使用决策树、支持向量机、卷积神经网络算法构建的模型，在对肺部疾病包括尘肺病诊断方面都取得了比较理想的结果。人工智能筛查或辅助诊断尘肺病，对减少读片差异和提高诊断准确性将发挥积极的作用。但建立人工智能模型需要使用胸片样本，而且要对胸片进行准确标记。
如采用传统方法，使用大量胸片样本，个人或集体对尘肺病胸片读片标注，由于受主观因素的影响，仍存在30～40％的较大差异，严重影响人工智能筛查或诊断尘肺病准确性，且成本效率不合理。
技术内容
本发明的目的是，提供一种采用统计功效估计方法遴选样本和并联与串联相结合读片标注方法，建立人工智能筛查尘肺病系统，通过验证获得理想效果。使人工智能在接触粉尘工人职业健康检查筛查和尘肺病辅助诊断中，提高筛查和诊断准确率，以及提高工作效率。
本发明的目的可以通过以下技术方案实现，一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法；人工智能系统包括计算机硬件系统及控制软件系统，计算机硬件系统包括中央处理器、存储器、输入输出控制系统；其所述方法包括以下步骤：
1)、基础数据库的建立：
(1)采用统计功效估计方法遴选DR(数字化X线摄影)胸片：选取DR正常胸片、壹期尘肺病胸片、贰期尘肺病胸片、叁期尘肺病胸片若干份；
(2)依据《职业性尘肺病诊断》(GBZ70-2015)，对选取的上述胸片进行分区，按两侧肺野6个肺区范围，将左肺野及右肺野，分别由肺尖至膈顶的垂直距离，用等分点的水平线把每侧肺野各分为上、中、下三个肺区，即将胸片左右肺野共分为6个区域；
(3)依据《职业性尘肺病诊断》(GBZ70-2015)，并对采用统计功效估计方法遴选的胸片，由人工进行每一区域的特征标注，具体特征标注为：
A、小阴影形态和大小：记录小阴影的形态和大小，分别用p、q、r、s、t、u标记或组合标注，如p/p、p/q、q/q、q/p、p/s、s/s、s/t、t/t、t/s等；
B、密集度：按三大级及十二小级分级标注，表示如下：0/-、0/0、0/1为0级；1/0、1/1、1/2为1级；2/1、2/2、2/3为2级；3/2、3/3、3/+为3级；
C、总体密集度：以4大级分级标注，表示如下：0、1、2、3级；
D、大阴影：用4表示，并标注所处肺区，同时在胸片上用红色线条画图标注；
E、小阴影聚集：用9表示，并标注所处肺区，同时在胸片上用绿色线条画图标注；
F、分布范围：将胸片中具有的上述特征，分别在两肺野相应肺区标注，以表示分布范围；
(4)人工标注：
采用并联与串联相结合方法，即双盲与集体读片方法读片标注，以 提高标准片的精准性；
A、将人工分为A、B两组，每组2人以上，各组分别独立读片，并用上述特征对每个肺区进行标注，诊断为尘肺及分期结果一致的胸片作为标准片；诊断结果不一致的胸片，随机再次分配给各组读片，诊断为尘肺及分期结果一致的胸片作为标准片；
B、对于两组读片结果不一致胸片，A组与B组共同读片，或另选同等条件的人工，即C组读片，诊断为尘肺及分期结果一致的作为标准片，认定未超过2/3一致的胸片剔除，每次读片间隔2周以上；
C、将标准片按分期结果分别存入计算机正常胸片标准数据库、壹期胸片标准数据库、贰期胸片标准数据库、叁期胸片标准数据库；
2)、计算与建模
采用深度学习方法之一的卷积神经网络法，对DR胸片DICOM图像进行学习训练，包括正常胸片标准数据库、壹期胸片标准数据库、贰期胸片标准数据库、叁期胸片标准数据库中的特征标注，建立人工智能筛查尘肺病初级模型；
由于原规格胸片分辨率较低，首先使胸片规格调整变小，对胸片两肺野进行抠影，并分为6个区域，以提高胸片分辨率，胸片重新调整至原规格，便于计算机对图像的识别；
模型共包含3个组件：输入层、隐藏层和输出层；其中，输入层用于接受DR胸片DICOM图像的输入,并传递给隐藏层，最终将信号传递给输出层；最后得出尘肺病及分期的预测结果；
3)、对人工智能模型验证：
(1)计算机录入用于验证的胸片，计算识别该胸片的特征，小阴影形态和大小、密集度、总体密集度、分布范围、大阴影、小阴影聚集，并分别存入待查胸片数据库；
(2)计算机将待查胸片数据库中该胸片的特征，输入筛查尘肺病初级模型，经过计算得出结果，生成尘肺病及分期的诊断报告；
(3)筛查尘肺病初级模型计算运行结束后，将该待查胸片、诊断结果存入诊断数据库；
(4)由两人以上的人工对诊断数据库中的疑似尘肺病的胸片读片，将读片最终诊断与计算机输出诊断结果对比进行验证，分别统计分析：阴阳准确率(ACC scr)、灵感度(SEN)、特异度(SPE)、分期准确率(ACC sta)；具体的计算为：
A、指标为：真阴性(TN)、伪阴性(FN)、伪阳性(FP)、真阳性(TP)；
B、计算公式为：
Figure PCTCN2020099674-appb-000001
Figure PCTCN2020099674-appb-000002
Figure PCTCN2020099674-appb-000003
Figure PCTCN2020099674-appb-000004
Figure PCTCN2020099674-appb-000005
Figure PCTCN2020099674-appb-000006
Figure PCTCN2020099674-appb-000007
(5)经上述统计分析，筛选尘肺病准确率达到97％或以上，分期准确率达到80％或以上，灵敏度0.98，特异度0.97，验证结果显示所建筛查尘肺病模型效果良好。
本发明可准确和高效的对尘肺病胸片进行标注，并用于人工智能筛查尘肺模型的建立，使人工智能在接触粉尘工人职业健康检查筛查和尘肺病辅助诊断中，提高筛查和诊断准确率、提高工作效率。
具体实施方式
一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法；人工智能系统包括计算机硬件系统及控制软件系统，计算机硬件系统包括中央处理器、存储器、输入输出控制系统；其特征在于所述方法包括以下步骤：
1、基础数据库的建立
(1)胸片遴选数量：本发明中胸片的遴选，是基于统计功效估计方式来确定遴选胸片张数的，遴选胸片以下统计功效是基于最小ROC曲线(接收者操作特征曲线)下面积(AUC)0.85来计算的。ROC曲线，即接收者操作特征曲线，是一种坐标图式的分析工具，用于选择最佳分类模型、舍弃次佳分类模型。AUC(Area under the Curve of ROC)，即ROC 曲线下方的面积。
对于N 1例正常胸片对比样本和N 2例尘肺壹期胸片样本，计算的统计功效为0.89。如果真实的AUC是0.90，那么模型更够准确的区分正常胸片样本和尘肺壹期胸片样本的AUC显著高于0.85(显著水平为0.05)。
对于N 2例尘肺壹期胸片样本和N 3例尘肺贰期胸片样本，计算的统计功效为0.90。如果真实的AUC是0.88，那么模型更够准确的区分正常胸片样本和尘肺壹期胸片样本的AUC显著高于0.85(显著水平为0.05)。
对于N 3例尘肺贰期胸片样本和N 4例尘肺叁期胸片样本，计算的统计功效为0.77。如果真实的AUC是0.88，那么模型更够准确的区分正常胸片样本和尘肺壹期胸片样本的AUC显著高于0.85(显著水平为0.05)。
根据计算结果，收集数字化X射线(Digital Radiography，DR)胸片DICOM原图数据共N例，胸片质量3级以上(不含3级)，其中壹期胸片、贰期胸片各N 2、N 3例，叁期胸片N 4例，健康人DR胸片N 1例，完全可以满足人工智能筛查尘肺病系统的需要。
(2)标注特征
1)依据《职业性尘肺病诊断》(GBZ70-2015)，对人工选取的上述胸片进行尘肺病胸片的分区，即将胸片左右肺野分为6个区域；
2)依据《职业性尘肺病诊断》(GBZ70-2015)，对人工选取的上述胸片由人工进行每一区域的特征标注，具体特征标注为；
a.小阴影形态和大小：记录小阴影的形态和大小，分别用p、q、r、s、t、u标记或组合标注，如p/p、p/q、q/q、q/p、p/s、s/s、s/t、t/t、t/s等；
b.密集度：按三大级及十二小级分级标注，表示如下：0/-、0/0、0/1为0级；1/0、1/1、1/2为1级；2/1、2/2、2/3为2级；3/2、3/3、3/+为3级；
c.总体密集度：以4大级分级标注，表示如下：0、1、2、3级；
d.大阴影：用4表示，并标注所处肺区，同时在胸片上用红色线条画图标注；
e.小阴影聚集：用9表示，并标注所处肺区，同时在胸片上用绿色线条画图标注。
f.分布范围：将胸片中具有的上述特征分别在两肺野相应肺区标注，以表示分布范围。
(3)标注方法
1)将具有尘肺病诊断资格、高级职称、从事专业技术工作10年以上专家分为A、B两组，每组2人以上，各组分别独立读片，并用上述方法对每个肺区进行标注，诊断为尘肺及分期结果一致的胸片作为标准片；诊断结果不一致的胸片，随机再次分配给各组读片，诊断为尘肺及分期结果一致的胸片作为标准片；
2)对于两组读片结果不一致胸片，A组与B组共同读片，或另选同等条件的人工，即C组读片，诊断为尘肺及分期结果一致的作为标准片，认定未超过2/3一致的胸片剔除，每次读片间隔2周以上；
3)将标准片按分期结果分别存入计算机正常胸片标准数据库、壹期胸片标准数据库、贰期胸片标准数据库、叁期胸片标准数据库。
2、计算与建模：
采用深度学习方法之一的卷积神经网络，分别对正常胸片标准数据库、壹期胸片标准数据库、贰期胸片标准数据库、叁期胸片标准数据库中的胸片的特征标注，包括小阴影形态和大小、密集度、总体密集度、分布范围、大阴影、小阴影聚集等，进行学习训练，并建立人工智能筛查尘肺病模型。所述模型共包含3个组件：输入层、隐藏层和输出层；其中，输入层用于接受原始DICOM图像的输入,并传递给隐藏层。隐藏层可设置不同层数的神经元，或采用不同的排布结构，并最终将信号传递给输出层。输出层汇总隐藏层最后一层的信号，得出尘肺胸片分期的预测结果。
3、对尘肺病模型验证
(1)计算机录入用于验证的胸片，存入待查胸片数据库，计算机识别该胸片的特征，小阴影形态和大小、密集度、总体密集度、分布范围、大阴影、小阴影聚集，并分别存入待查胸片影像特征数据库；
(2)计算机将待查胸片影像特征数据库中该胸片的特征，输入筛查尘肺病初期模型，经计算，并给出诊断结果，生成具有上述数据的诊断报告；
(3)计算机将该待查胸片、诊断结果存入诊疗数据库。
(4)由具有尘肺病诊断资格、高级职称、从事专业技术工作10年以上专家，集体对计算得出的疑似尘肺病结果的胸片读片，将读片最终诊断与计算机输出诊断结果对比，进行验证，并统计分析：阴阳准确率(ACC scr)、灵感度(SEN)、特异度(SPE)、分期准确率(ACC sta)；具体的计算为：
1)指标为：真阴性(TN)、伪阴性(FN)、伪阳性(FP)、真阳性(TP)；
2)计算公式为：
Figure PCTCN2020099674-appb-000008
Figure PCTCN2020099674-appb-000009
Figure PCTCN2020099674-appb-000010
Figure PCTCN2020099674-appb-000011
Figure PCTCN2020099674-appb-000012
Figure PCTCN2020099674-appb-000013
Figure PCTCN2020099674-appb-000014
(5)经上述统计分析，筛选准确率达到97％或以上，分期准确率达到80％或以上，灵敏度0.98，特异度0.97，验证结果：所建筛查尘肺病模型效果良好。

Claims (1)

  1. 一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法；人工智能系统包括计算机硬件系统及控制软件系统，计算机硬件系统包括中央处理器、存储器、输入输出控制系统；其特征在于所述方法包括以下步骤：
    1)、基础数据库的建立：
    (1)采用统计功效估计方法遴选DR(数字化X线摄影)胸片：选取DR正常胸片、壹期尘肺病胸片、贰期尘肺病胸片、叁期尘肺病胸片若干份；
    (2)依据《职业性尘肺病诊断》(GBZ70-2015)，对选取的上述胸片进行分区，按两侧肺野6个肺区范围，将左肺野及右肺野，分别由肺尖至膈顶的垂直距离，用等分点的水平线把每侧肺野各分为上、中、下三个肺区，即将胸片左右肺野共分为6个区域；
    (3)依据《职业性尘肺病诊断》(GBZ70-2015)，并对采用统计功效估计方法遴选的胸片，由人工进行每一区域的特征标注，具体特征标注为：
    A、小阴影形态和大小：记录小阴影的形态和大小，分别用p、q、r、s、t、u标记或组合标注，如p/p、p/q、q/q、q/p、p/s、s/s、s/t、t/t、t/s等；
    B、密集度：按三大级及十二小级分级标注，表示如下：0/-、0/0、0/1为0级；1/0、1/1、1/2为1级；2/1、2/2、2/3为2级；3/2、3/3、3/+为3级；
    C、总体密集度：以4大级分级标注，表示如下：0、1、2、3级；
    D、大阴影：用4表示，并标注所处肺区，同时在胸片上用红色线条画图标注；
    E、小阴影聚集：用9表示，并标注所处肺区，同时在胸片上用绿色线条画图标注；
    F、分布范围：将胸片中具有的上述特征，分别在两肺野相应肺区标注，以表示分布范围；
    (4)人工标注：
    采用并联与串联相结合方法，即双盲与集体读片方法读片标注，以提高标准片的精准性；
    A、将人工分为A、B两组，每组2人以上，各组分别独立读片，并用上述特征对每个肺区进行标注，诊断为尘肺及分期结果一致的胸片作为标准片；诊断结果不一致的胸片，随机再次分配给各组读片，诊断为尘肺及分期结果一致的胸片作为标准片；
    B、对于两组读片结果不一致胸片，A组与B组共同读片，或另选同等条件的人工，即C组读片，诊断为尘肺及分期结果一致的作为标准片，认定未超过2/3一致的胸片剔除，每次读片间隔2周以上；
    C、将标准片按分期结果分别存入计算机正常胸片标准数据库、壹期胸片标准数据库、贰期胸片标准数据库、叁期胸片标准数据库；
    2)、计算与建模
    采用深度学习方法之一的卷积神经网络法，对DR胸片DICOM图像进行学习训练，包括正常胸片标准数据库、壹期胸片标准数据库、贰期胸片标准数据库、叁期胸片标准数据库中的特征标注，建立人工智能筛查 尘肺病初级模型；
    由于原规格胸片分辨率较低，首先使胸片规格调整变小，对胸片两肺野进行抠影，并分为6个区域，以提高胸片分辨率，胸片重新调整至原规格，便于计算机对图像的识别；
    模型共包含3个组件：输入层、隐藏层和输出层；其中，输入层用于接受DR胸片DICOM图像的输入,并传递给隐藏层，最终将信号传递给输出层；最后得出尘肺病及分期的预测结果；
    3)、对人工智能模型验证：
    (1)计算机录入用于验证的胸片，计算识别该胸片的特征，小阴影形态和大小、密集度、总体密集度、分布范围、大阴影、小阴影聚集，并分别存入待查胸片数据库；
    (2)计算机将待查胸片数据库中该胸片的特征，输入筛查尘肺病初级模型，经过计算得出结果，生成尘肺病及分期的诊断报告；
    (3)筛查尘肺病初级模型计算运行结束后，将该待查胸片、诊断结果存入诊断数据库；
    (4)由两人以上的人工对诊断数据库中的疑似尘肺病的胸片读片，将读片最终诊断与计算机输出诊断结果对比进行验证，分别统计分析：阴阳准确率(ACC scr)、灵感度(SEN)、特异度(SPE)、分期准确率(ACC sta)；具体的计算为：
    A、指标为：真阴性(TN)、伪阴性(FN)、伪阳性(FP)、真阳性(TP)；
    B、计算公式为：
    Figure PCTCN2020099674-appb-100001
    Figure PCTCN2020099674-appb-100002
    Figure PCTCN2020099674-appb-100003
    Figure PCTCN2020099674-appb-100004
    Figure PCTCN2020099674-appb-100005
    Figure PCTCN2020099674-appb-100006
    Figure PCTCN2020099674-appb-100007
    (5)经上述统计分析，筛选尘肺病准确率达到97％或以上，分期准确率达到80％或以上，灵敏度0.98，特异度0.97，验证结果显示所建筛查尘肺病模型效果良好。
PCT/CN2020/099674 2020-07-01 2020-07-01 一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法 WO2022000360A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/099674 WO2022000360A1 (zh) 2020-07-01 2020-07-01 一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/099674 WO2022000360A1 (zh) 2020-07-01 2020-07-01 一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法

Publications (1)

Publication Number Publication Date
WO2022000360A1 true WO2022000360A1 (zh) 2022-01-06

Family

ID=79317777

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/099674 WO2022000360A1 (zh) 2020-07-01 2020-07-01 一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法

Country Status (1)

Country Link
WO (1) WO2022000360A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105574871A (zh) * 2015-12-16 2016-05-11 深圳市智影医疗科技有限公司 在放射图像中检测肺部局部性病变的分割分类方法和系统
CN106971198A (zh) * 2017-03-03 2017-07-21 北京市计算中心 一种基于深度学习的尘肺病等级判定方法及系统
US20190171428A1 (en) * 2017-12-04 2019-06-06 Banjo, Inc. Automated model management methods
CN110680326A (zh) * 2019-10-11 2020-01-14 北京大学第三医院（北京大学第三临床医学院） 基于深度卷积神经网络的尘肺病鉴别及分级判定方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105574871A (zh) * 2015-12-16 2016-05-11 深圳市智影医疗科技有限公司 在放射图像中检测肺部局部性病变的分割分类方法和系统
CN106971198A (zh) * 2017-03-03 2017-07-21 北京市计算中心 一种基于深度学习的尘肺病等级判定方法及系统
US20190171428A1 (en) * 2017-12-04 2019-06-06 Banjo, Inc. Automated model management methods
CN110680326A (zh) * 2019-10-11 2020-01-14 北京大学第三医院（北京大学第三临床医学院） 基于深度卷积神经网络的尘肺病鉴别及分级判定方法

Similar Documents

Publication Publication Date Title
CN111816314B (zh) 一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法
CN107103187A (zh) 基于深度学习的肺结节检测分级与管理的方法及系统
CN109858540B (zh) 一种基于多模态融合的医学图像识别系统及方法
JP6807820B2 (ja) 画像検索装置、方法およびプログラム
CN108389201A (zh) 基于3d卷积神经网络与深度学习的肺结节良恶性分类方法
Li et al. Computer-aided diagnosis of COVID-19 CT scans based on spatiotemporal information fusion
Tobias et al. CNN-based deep learning model for chest X-ray health classification using tensorflow
CN111598868B (zh) 一种肺部超声图像识别方法和系统
Sathya et al. Transfer learning based automatic human identification using dental traits-an aid to forensic odontology
Dinesh et al. Machine Learning in the Detection of Oral Lesions With Clinical Intraoral Images
CN113397485A (zh) 一种基于深度学习的脊柱侧弯筛查方法
CN112690815A (zh) 基于肺部图像报告辅助诊断病变等级的系统和方法
CN113362289A (zh) 一种胸部dr图像上异物自动检测的方法
WO2022000360A1 (zh) 一种用于人工智能筛查尘肺病胸片遴选、标注及验证方法
CN112489022A (zh) 基于放射组学的covid-19快速特征提取系统及快速诊断系统及病程预测系统
CN115602327A (zh) 一种肺结节发生肺癌风险的预测模型的构建方法
US11941811B2 (en) Method for assessing cardiothoracic ratio and cardiothoracic ratio assessment system
CN116402756A (zh) 一种融合多层次特征的x光片肺部疾病筛查系统
CN114998203A (zh) 一种基于人工智能的职业性尘肺病精准诊断系统及方法
Wang et al. Intelligent image diagnosis of pneumoconiosis based on wavelet transform-derived texture features
Nneji et al. COVID-19 Identification Using Deep Capsule Network: A Perspective of Super-Resolution CNN on Low-Quality CXR Images
CN113344022A (zh) 一种基于深度学习的胸片检测方法
Liu et al. Combining ExtremeNet with Shape Constraints and Re-Discrimination to Detect Cells from CD56 Images
Wu et al. Attention-guided multiple instance learning for COPD identification: To combine the intensity and morphology
AU2021105838A4 (en) System for Detecting and Screening Pulmonary Nodules Based on Artificial Intelligence

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20942496

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20942496

Country of ref document: EP

Kind code of ref document: A1