| Objective: Pneumoconiosis is a severe occupational disease caused by long-term inhalation of high-concentration dust in the production environment.At present,there is no effective treatment.As a progressive disease,pneumoconiosis is always progressive,even after complete cessation of exposure.With the progression of the disease,patients will have pulmonary dysfunction,dyspnea and other symptoms.Severe patients gradually lose working ability and even die,causing huge losses and burdens to families and society.China has the largest number of cases of pneumoconiosis.In 2016,the political bureau of the central committee of the communist party of China reviewed and adopted the "Healthy China 2030" plan,which emphasized the prevention and early diagnosis and treatment of occupational diseases.In 2019,the National Health Commission of the People’s Republic of China and other departments jointly issued the "Action Plan for Pneumoconiosis Prevention and Control" to continuously strengthen the prevention and control of pneumoconiosis and the treatment and assistance of pneumoconiosis patients.Thus,our country’s pneumoconiosis prevention and control task is very arduous.Governments,enterprises and medical institutions have invested a lot of human,material and financial resources for a long time.However,as a key in the prevention and treatment of pneumoconiosis,the diagnosis of pneumoconiosis is extremely difficult.According to the current diagnostic criteria for pneumoconiosis in China,the diagnosis is mainly based on the chest radiograph findings of dust-exposed workers,made by qualified radiologists according to the shape,size,profusion and distribution of the lesion in the chest radiograph.However,as a kind of interstitial lung disease,the chest radiographs of pneumoconiosis show the characteristics of diversity and complexity,which makes the diagnostic method of manual reading have a strong subjectivity.The level and experience of radiologists will affect the accuracy of diagnosis,and there are inevitable differences in reading among radiologists.The dustexposed workers in the early stage of pneumoconiosis have no obvious clinical symptoms,and the chest radiograph findings are extremely similar to the normal X-rays.Therefore,this difference is particularly obvious in the diagnosis of non-pneumoconiosis and pneumoconiosis with stage I,which is easy to be missed and misdiagnosed.It can be seen that it is very difficult to carry out early screening for pneumoconiosis.On the other hand,the application of artificial intelligence(AI)technology to solve clinical problems has become one of the research hotspots,especially in the interpretation of medical images,which shows great potential.In the field of occupational health,the diagnosis of pneumoconiosis based on AI is a major research topic.However,due to the complexity of pneumoconiosis diagnosis,such research is still in its infancy.The main technical bottleneck restricting the application of AI-assisted diagnosis of pneumoconiosis is how to effectively identify non-pneumoconiosis and stage I pneumoconiosis.At present,there are few studies on intelligent diagnosis of early pneumoconiosis.In particular,there are no studies on the identification of chest radiographs for pneumoconiosis with subcategory 0/1 pneumoconiosis.Therefore,it is very urgent to conduct in-depth research on the AI-based diagnosis of pneumoconiosis and propose an integrated diagnostic scheme,which is necessary to carry out large-scale pneumoconiosis screening in clinics.First,it can greatly improve the reading efficiency of radiologists,automate tedious and repetitive work,and reduce the radiologists’ burden.Second,the skilled services can be expanded.The use of AI scheme to assist the diagnosis of radiologists can improve the consistency and repeatability of the diagnosis between different seniority,regions,hospitals and radiologists,effectively solve the difference in the diagnosis of pneumoconiosis,and reduce the incidence of missed diagnosis and misdiagnosis.In addition,it will help solve the shortage of specialists in grass-roots units,accelerate the dissemination of advanced diagnostic and therapeutic skills to primary hospitals and reduce medical costs.Finally,dust-exposed workers in the early stage of pneumoconiosis can receive timely and effective intervention,delay the development of pneumoconiosis and improve the prognosis.The research on AI-based diagnosis of pneumoconiosis will eventually serve as the prevention of public health and has important practical significance and social benefits in the early warning of high-risk occupational groups and secondary prevention of susceptible groups.In recent years,the research methods of pneumoconiosis diagnosis based on artificial intelligence have developed from traditional machine learning(ML)to deep learning(DL).Compared with traditional machine learning methods that require manual feature extraction,deep learning algorithms represented by convolutional neural network(CNN)can automatically extract and analyze features such as lung texture in chest radiographs during model training,to realize automatic diagnosis.It has become the mainstream scheme for diagnosing pneumoconiosis based on AI.In the DL scheme,there are some specific problems such as expanding the sample size and enhancing the image quality of chest radiographs,segmenting the lung region in chest radiographs and effectively improving the prediction performance of deep learning models,which is necessary to conduct in-depth study.In summary,the overall aim of this study is to implement an early prediction model and screening scheme for pneumoconiosis based on DL.This scheme can quickly and effectively screen out the early stage of pneumoconiosis in a large number of DRs of dustexposed workers,especially those with subcategory 0/1 pneumoconiosis.The study first analyzes the specific effects of image quality enhancement and data enhancement strategies on the performance of the early pneumoconiosis recognition model based on the original chest radiographs of dust-exposed workers.Secondly,the lung region image set and the lung target boxes image set were used to train the model,and the prediction performance of the model was compared and analyzed.Finally,based on the higher-quality image data provided by the above two steps,the ensemble learning strategy was applied to construct the final early pneumoconiosis recognition model and integrated screening scheme.Methods: First,an original dataset of pneumoconiosis chest radiographs from dust-exposed workers was created,the original radiographs in the dataset were pre-processed for image quality enhancement,and the sample size of the training set was expanded using a combination of offline and online data enhancement.Secondly,a YOLOv2 lung target recognition network was trained to extract lung target boxes from the chest radiographs.Thirdly,seven CNNs were trained for identifying early pneumoconiosis,and an ensemble learning(EL)model based on the soft voting strategy was built.Finally,the performance of the EL model was evaluated on the test set.The specific research methods are as follows.1.A total of 1634 digital radiographs(DR)with baseline data of occupational workers from two hospitals between July 2015 and August 2021 were collected.A total of 679 normal DRs(profusion of small opacities of 0/-subcategory and 0/0 subcategory)and 690 abnormal DRs(profusion of small opacities of 0/1 subcategory,stage 1 pneumoconiosis,stage 2 pneumoconiosis and stage 3 pneumoconiosis)were used as the original chest radiographs data set according to the established inclusion and exclusion criteria.2.Based on the original chest radiograph dataset,four image pre-processing procedures were applied respectively to generate four image pre-processing datasets,including Contrast Limited Adaptive Histogram Equalization(CLAHE),Histogram Equalization(HE),the addition of Gaussian noise and Gamma correction.These image pre-processing datasets were combined with the original image dataset to create the augmentation dataset,which expanded the training set.3.A lightweight convolutional neural network,Shuffle Net,was used to train the image augmentation dataset to create the early pneumoconiosis identification model.The model performance was evaluated by the receiver operating characteristic curve(ROC)and the area under curve(AUC),Accuracy,Sensitivity,Specificity and Youden index.4.A total of 485 chest radiographs with the corresponding mask images from four public chest radiograph datasets were collected and were used to create and test a lung region segmentation model and a lung target box recognition model.A U-Net-based lung region segmentation model was trained to segment and extract the lung region from the chest radiograph images.The segmentation model was evaluated using Accuracy,Precision,Recall,Intersection-over-Union(Io U)and Dice Similarity Coefficient(DSC).A YOLOv2-based lung target box recognition model was also trained to identify lung targets from chest radiograph images and generate rectangular lung box regions.The model was also evaluated using Precision,Recall,F1 score and Precision-Recall curve(P-R curve).5.Lung region segmentation and lung target box identification were performed using the U-Net lung region segmentation model and the YOLOv2 lung target box recognition model on the image augmentation dataset to generate the lung region image dataset and the lung target box image dataset,respectively.The Shuffle Net-based early pneumoconiosis recognition models were trained on the two datasets separately.The performance of the two models was evaluated using AUC values,Accuracy,Sensitivity,Specificity and Youden index.6.Six CNNs(Inception-V3,Xception,Res Net101,Dense Net,Mobile Net and Efficient Net)were trained based on the lung target box image dataset with the previously completed Shuffle Net model made up the seven base learners for ensemble learning.The prediction results of the seven base learners were integrated using a soft voting method to build the final ensemble learning model for early pneumoconiosis identification.The performance of the ensemble learning model was also evaluated using AUC values,Accuracy,Sensitivity,Specificity and Youden index.Results: 1.The Shuffle Net model for early-stage pneumoconiosis identification trained on the image augmentation dataset obtained the AUC value of 0.891(95% CI: 0.826-0.938,P<0.01),the accuracy of 85.2%(95% CI: 78.1%-90.7%,P<0.01),the sensitivity of 88.4%(95% CI: 78.4%-94.9%,P<0.01),the specificity of 81.8%(95% CI: 70.4%-90.2%,P<0.01)and the Youden index of 0.7022 on the test set.2.The U-Net lung region segmentation model trained on the public chest radiograph dataset generated the overall accuracy of 94.31%,the precision of 97.93%,the recall of81.13%,the average cross-merge ratio of 86.21% and the Dice similarity coefficient of88.74% on the test set.3.The YOLOv2 lung target box recognition model trained on the public chest radiograph dataset generated the accuracy of 99.49%,the recall of 97.00% and the F1 score of0.9823 on the test set when the lung target determination threshold was 0.5.4.The Shuffle Net model for early-stage pneumoconiosis identification trained on lung region images generated the AUC value of 0.914(95% CI: 0.853-0.955,P<0.01),the accuracy of 83.0%(95% CI: 75.5%-88.9%,P<0.01),the sensitivity of 82.6%(95% CI:71.6%-90.7%,P<0.01),the specificity of 83.3%(95% CI: 72.1%-91.4%,P< 0.01)and the Youden index of 0.6594 on the test set.5.The Shuffle Net model for early-stage pneumoconiosis identification trained on lung target box images obtained the AUC value of 0.919(95%CI: 0.859-0.959,P<0.01),the accuracy of 85.2%(95%CI: 78.1%-90.7%,P<0.01),the sensitivity of 88.4%(95%CI:78.4%-94.9%,P<0.01),the specificity of 81.8%(95%CI: 70.4%-90.2%,P<0.01)and the Youden index of 0.7022 on the test set.6.The Iception V3 model for early-stage pneumoconiosis identification trained on lung target box images obtained the AUC value of 0.918(95% CI: 0.858-0.958,P<0.01),the accuracy of 86.7%(95% CI: 79.8%-91.9%,P<0.01),the sensitivity of 82.6%(95% CI:71.6%-90.7%,P<0.01),the specificity of 90.9%(95% CI: 81.3%-96.6%,P< 0.01)and the Youden index of 0.7352 on the test set.7.The Xception model for early-stage pneumoconiosis identification trained on lung target box images obtained the AUC value of 0.924(95% CI: 0.865-0.962,P<0.01),the accuracy of 89.6%(95% CI: 83.2%-94.2%,P<0.01),the sensitivity of 88.4%(95% CI:78.4%-94.9%,P<0.01),the specificity of 90.9%(95% CI: 81.3%-96.6%,P<0.01)and the Youden index of 0.7931 on the test set.8.The Res Net101 model for early-stage pneumoconiosis identification trained on lung target box images obtained the AUC value of 0.892(95% CI: 0.828-0.939,P<0.01),the accuracy of 83.0%(95% CI: 75.5%-88.9%,P<0.01),the sensitivity of 82.6%(95% CI:71.6%-90.7%,P<0.01),the specificity of 83.3%(95% CI: 72.1%-91.4%,P< 0.01)and the Youden index of 0.6594 on the test set.9.The Dense Net model for early-stage pneumoconiosis identification trained on lung target box images obtained the AUC value of 0.894(95% CI: 0.829-0.940,P<0.01),the accuracy of 83.0%(95% CI: 75.5%-88.9%,P<0.01),the sensitivity of 82.6%(95% CI:71.6%-90.7%,P<0.01),the specificity of 83.3%(95% CI: 72.1%-91.4%,P<0.01)and the Youden index of 0.6594 on the test set.10.The Mobile Net model for early-stage pneumoconiosis identification trained on lung target box images obtained the AUC value of 0.921(95% CI: 0.861-0.960,P<0.01),the accuracy of 87.4%(95% CI: 80.6%-92.5%,P<0.01),the sensitivity of 85.5%(95% CI:75.0%-92.8%,P<0.01),the specificity of 89.4%(95% CI: 79.4%-95.6%,P< 0.01)and the Youden index of 0.7490 on the test set.11.The Efficient Net model for early-stage pneumoconiosis identification trained on lung target box images obtained the AUC value of 0.894(95% CI: 0.830-0.941,P<0.01),the accuracy of 82.2%(95% CI: 74.7%-88.3%,P<0.01),the sensitivity of 84.1%(95% CI:73.3%-91.8%,P<0.01),the specificity of 80.3%(95% CI: 68.7%-89.1%,P< 0.01)and the Youden index of 0.6436 on the test set.12.The ensemble learning model based on hard voting for early-stage pneumoconiosis identification obtained the AUC value of 0.934(95% CI: 0.878-0.969,P<0.01),the accuracy of 93.3%(95% CI: 87.7%-96.9%,P<0.01),the sensitivity of 91.3%(95% CI:82.0%-96.7%,P<0.01),the specificity of 95.5%(95% CI: 87.3%-99.1%,P<0.01)and the Youden index of 0.8676 on the test set.13.The ensemble learning model based on soft voting for early-stage pneumoconiosis identification obtained the AUC value of 0.959(95% CI: 0.911-0.986,P<0.01),the accuracy of 89.6%(95% CI: 83.2%-94.2%,P<0.01),the sensitivity of 82.6%(95% CI: 71.6%-90.7%,P<0.01),the specificity of 97.0%(95% CI: 89.5%-99.6%,P<0.01)and the Youden index of 0.7958 on the test set.Conclusions: 1.The combination of offline and online data augmentation techniques can effectively expand the training sample and improve the predictive performance of the early pneumoconiosis identification model.2.The use of lung target box images can further improve the performance of early pneumoconiosis identification models compared to the original chest radiographs.3.The early pneumoconiosis identification model and integrated screening scheme constructed by deep ensemble learning strategy based on soft voting have better comprehensive performance,which can quickly and effectively screen early and more severe pneumoconiosis,especially the chest radiographs with subcategory 0/1. |
PDF Full Text Request