Diagnosis Of Pneumoconiosis At DR By Deep Convolutional Neural Networks

Objectives Based on deep convolutional neural network,three deep learning models were constructed to diagnose pneumoconiosis abnormalities.The diagnostic efficiency of each model was evaluated and the best model was selected as the artificial intelligence diagnostic method for pneumoconiosis.Methods We cooperated with 7 hospitals in Beijing,Henan,Ningxia,Jiangxi and other provinces and cities,and collected the DR chest radiographs between June 2017 to December 2020,control the quality of DR chest radiographs also.DR chest radiographs with positive diagnosis of pneumoconiosis were in the positive group,while those without pneumoconiosis were in the negative group.Two physicians diagnosing pneumoconiosis were selected for preliminary screening of the collected chest radiographs,and Grade III and IV radiographs were excluded.We examined the diagnostic experts for pneumoconiosis,chest radiographs shall be labeled by qualified experts,and experts are constantly assessed for consistency in the labeling process,which is based on expectationmaximization algorithm.The labeled data were cleaned,archived and image preprocessed,and the preprocessed data were incorporated into the training set and verification set.We built three deep learning models,TMNet,Res Net-50 and Res Ne Xt-50,and the models were pre-trained on a dataset of Inmage Net(a large visual database of more than 14 million images),the training was then transferred to the Chestx-Ray 14 dataset(a database of over 100,000 chest X-ray images containing 14 common chest diseases)and transfer learning to the pneumoconiosis data set at last.The model was evaluated by ten-fold crossvalidation method to get the optimal model.Finally 500 cases of DR chest radiographs that were not included in the training set and validation set were selected,and these selected DR chest radiographs were identified by five senior experts as the gold standard,named the test set.With the increase of the number of chest radiographs collected(a total of 3tests),the accuracy rate,F1 score,sensitivity,specificity,AUC and other indexes of the three models were obtained from the test results,and the efficiency of the three models was evaluated and compared..Results 1 Chest radiograph data showed that a total of 24,867 patients were included in the study,with 28.06% in the positive group and 71.94% in the negative group.In the positive group,the ratio of coal workers’ pneumoconiosis and silicosis was 90.82%,asbestosis was 5.74%,and the rest was other pneumoconiosis.In the quality of DR chest radiographs,the first grade radiographs accounted for 25.19% and the second grade radiographs 74.81%.The proportion of stage I,II and III pneumoconiosis in positive group was 56.51%,25.09% and 18.40% respectively.A total of 500 cases DR chest radiographs were collected,including 298(59.60%)cases in the positive group and 202(40.40%)cases in the negative group.In the positive group,201(67.45%)cases of stage Ⅰ pneumoconiosis,68(22.82%)cases of stage II pneumoconiosis and 29(9.73%)cases of stage III pneumoconiosis were found.In terms of the quality of DR chest radiographs,the number of first grade and second grade radiographs accounted for 20.60% and 79.40% respectively.A total of 312 cases of abnormal pulmonary conditions,such as pneumothorax and tuberculosis,were collected on DR chest radiographs,among which 93 cases were active tuberculosis.2 The results of expert labling showed that the consistency rate of with or without pneumoconiosis was above 88%,and the consistency rate of pneumoconiosis staging ranged from 84.68% to 93.66%.3 The data volume of the first round of model test was 8,341 cases,and the diagnostic efficiency of TMNet was better than the other two models.The accuracy,sensitivity and specificity of TMNet model were 85.00%,85.23%and 84.65%,and the AUC value was 0.887.There were statistically significant differences(P<0.05)in AUC values between model TMNet and model Res Net-50,and between model TMNet and model Res Ne Xt-50.4 The data volume of the second round of model test was14,319 cases.The accuracy,sensitivity and specificity of the TMNet model were improved to 92.60%,94.30% and 90.10% respectively,and the AUC value was 0.948.The difference of AUC between the TMNet model and the other two models was statistically significant(P<0.05).5 The volume of model test data in the third round was 24,867 cases.Similar to the previous two rounds,the AUC difference between the TMNet model and the other two models was still statistically significant(P<0.05),the diagnostic accuracy,sensitivity,specificity and AUC of TMNet were 95.20%,99.66%,88.61% and 0.987.6After the comparison of the three models,the performance of the models was improved with the increase of the data volume,and the difference between the two AUC values before and after the models was statistically significant(P<0.05).Conclusions All the three convolutional neural network models can diagnose pneumoconiosis,and their performance will be improved with the increase of data volume.By comprehensive comparison,TMNet model has the best diagnostic performance,which can reach 95% accuracy and extremely high sensitivity and specificity,and can most accurately diagnose of pneumoconiosis.Figure 25;Table 15;Reference 97...