| Pneumoconiosis is the number one occupational disease with the highest number of incidences and most common in China.Early clarification of pneumoconiosis staging and early selection of targeted and correct treatment plan can significantly improve patients’ quality of life and survival rate,which has important clinical value.According to the diagnostic criteria of GBZ 70-2015 Occupational Pneumoconiosis Diagnosis,digital radiograph(DR)is the main examination method for pneumoconiosis diagnosis,and the core of pneumoconiosis diagnosis is the correct interpretation of pneumoconiosis X-ray chest film.In clinical work,radiologists mainly analyze the image features such as the size,shape,density,distribution of lung areas and the presence or absence of large shadows,and finally give the correct pneumoconiosis staging by combining the patient’s history of receiving dust,relevant laboratory tests and clinical manifestations.However,the diagnostic process of pneumoconiosis is complex,and there are differences in reading before and after the reviewer himself or herself and between the reviewers,especially inexperienced junior physicians,who find it difficult to accurately identify important imaging signs and give accurate staging,so how to improve the stability and accuracy of the reviewers is an urgent problem to be solved.Based on this,this study retrospectively collected X-ray images of different stages of pneumoconiosis screening and analyzed the accuracy and stability of clinical pneumoconiosis staging assessment with reference to the diagnostic criteria of GBZ 70-2015 Occupational Pneumoconiosis Diagnosis to provide reference for the necessity of constructing a deep learning based end-to-end automatic pneumoconiosis classification model.[Methods]1.Clinical informationA total of 980 cases of pneumoconiosis screening in The twelfth people’s Hospital of Guangzhou from 2016 to 2019 were collected.The quality of chest X-ray conforms to the first-grade film in Appendix C of GBZ 70-2015.2.Operation methodSiemens Axiom Aristos X-ray double lithography digital imaging system(DR),Kodak DV medical infrared laser film and Kodak 8900 dry laser washer were used for DR chest X-ray examination.After the patient fully inhaled,the standard chest radiographs were taken in automatic exposure mode.The residual neural network and dense convolution neural network(ResNet50,ResNet10,DenseNet)are selected as the framework of the classification network,and the model is constructed.The optimization function(Adam),activation function(ReLU)and loss function(Categorical-Crossentropy)are used to apply them to the data set,and the training starts from scratch.After the model training is completed,the model is applied to the new image,and finally the classification performance of the model is evaluated.3.Evaluation methodThe first part:the imaging staging of pneumoconiosis was evaluated by three pneumoconiosis imaging diagnostic physicians with different seniority(low,medium and high,with diagnostic experience of more than 5,10 and 15 years respectively).The initial assessment results were collected from 3 physicians of different seniority(RI R2,and R3);the classification was extracted from the original report of each case and labeled as Report;The staging agreed by the three physicians was the relative gold standard,which was recorded as Major.The second part:using precision,recall and comprehensive classification are used as the classification indexes of a single model,macro-average and micro-average are used to evaluate the overall classification effect of the model,and the receiver operation characteristic curve(receiver operating characteristic,ROC)analysis and the area under ROC curve(area under the curve,AUC)are used to measure the classification performance of the model.4.Statistical methodStatistical analysis software uses SPSS 20.0,uses Cohen’s Kappa test to evaluate the consistency of classification,including the consistency between the reader and the relative gold standard,the consistency of the original report and the relative gold standard,and uses the accuracy rate is used to evaluate the accuracy of each viewer’s staging.The evaluation criteria of Kappa value refer to Altman guidelines.Kappa value<0.4,0.4 ≤ Kappa<0.75 as medium consistency,Kappa>0.75 as better consistency,P<0.05.The difference is statistically significant.The performance of the model classifier for pneumoconiosis staging was evaluated by receiver operating characteristic curve(ROC)analysis and area under ROC curve(AUC).[Results]1.Different readers have different understanding and mastery of the standard"GBZ 70-2015 diagnosis of Occupational Pneumoconiosis".The original report Kappa is 0.583 and the consistency is moderate;R1(junior physician)and relative gold standard Kappa value is 0.745,moderate consistency;R2,R3(middle and senior physicians)and relative gold standard Kappa are 0.836,0.896,respectively,the consistency is good.Two classification methods:the Kappa of the original report was 0.599,the consistency was moderate,the Kappa of R1 and the relative gold standard was 0.714,the Kappa of R2,R3 and the relative gold standard was 0.819,0.897 respectively,the consistency was good,the above results were all P<0.001.The differences were mainly concentrated in normal/stage Ⅰ and stage Ⅰ/Ⅱ.Comparing the coincidence rate of image features with great differences between different periods,the image features are mainly reflected in the small shadow shape,overall density and lung area distribution of stage Ⅰ and Ⅱ,in which the lowest coincidence rate is the small shadow shape,which is mainly reflected in the judgment of low and middle-aged doctors in stage Ⅰ and Ⅱ,with a coincidence rate of 43.30%and 54.10%,respectively.2.In the aspect of model construction of end-to-end automatic pneumoconiosis classification based on deep learning,the macro accuracy,macro recall,macro comprehensive classification and micro average of ResNet50 single model are 0.83,0.82,0.82,0.82 and 0.82 respectively.The macro accuracy,macro comprehensive classification rate and micro average of single model ResNet101 are 0.79,0.78,0.77 and 0.77,respectively.The macro precision rate,macro recall rate,macro comprehensive classification rate and micro average of DenseNet single model were 0.81,0.80,0.80,0.80 respectively.ResNet50 model showed that the precision rates of 0,1,Ⅱ and Ⅲ were 0.91,0.84,0.86,0.72,the recall rates were 1.0,0.80,0.60,0.90,and the comprehensive classification rates were 0.95,0.82,0.71 and 0.80,respectively.The index of the overall efficiency of the model:the micro-average and macro-average of the area under the operating characteristic curve of the subjects were 0.93,0.94,respectively.the ResNet50 model showed that the precision rates of 0,Ⅰ,Ⅱ and Ⅲ were 0.91,0.84,0.60,0.90 and 0.95,0.82,0.71 and 0.80,respectively.ResnetlOl model showed that the precision rates of each classification in 0,Ⅰ,Ⅱ andⅢ were 0.95,0.71,0.64 and 0.86,the recall rates were 0.95,0.85,0.70,0.60 respectively,and the comprehensive classification rates were 0.95,0.77,0.67 and 0.71 respectively.The index of the overall efficiency of the model:the(AUC)micro-average and macro-average of the area under the operating characteristic curve of the subjects were 0.92,0.94 respectively.DenseNet model showed that the precision rates of each classification in periods 0,Ⅰ,Ⅱ and Ⅲ were 1.0,0.75,0.67 and 0.83,the recall rates were 0.90,0.75,0.80,0.75,and the comprehensive classification rates were 0.95,0.75,0.73 and 0.79,respectively.The index of the overall efficiency of the model:the(AUC)micro-average and macro-average of the area under the operating characteristic curve of the subjects were 0.94,0.95 respectively.Compared with the observation indexes of the above three models,the results show that the efficiency of DenseNet model is the best,and its classification effect in stage 0 and stage Ⅲ is better,while the classification effect in stage II needs to be optimized.[Conclusion]1.There is variability in the diagnostic staging categories of pneumoconiosis,as different reviewers have different mastery and proficiency in pneumoconiosis diagnostic criteria and perception of pneumoconiosis imaging features.2.The reliability of pneumoconiosis classification assessment in clinical work is generally limited,and there is a clinical need for a more consistent classification method,which can be improved by specifically developing a tool for automatic pneumoconiosis classification based on the standard "GBZ 70-2015 Diagnosis of Occupational Pneumoconiosis" to exclude the interference of subjective factors and improve the accuracy rate for more accurate interpretation of pneumoconiosis staging types.3.The end-to-end automatic pneumoconiosis classification method based on deep learning can effectively identify the basic image signs and distinguish four different types of pneumoconiosis from a certain amount of training data.On the whole,the macro-average and micro-average of ResNet50 and DenseNet were more than 0.80,indicating that the overall classification efficiency of the model was good.The area(AUC)micro-average and macro-average under the operating characteristic curve of ResNet50 subjects were 0.93,0.94 and 0.94 respectively.(AUC)micro-average and macro-average were 0.94 and 0.95 respectively.The overall classification efficiency of DenseNet model is the best,and it is a more suitable classification model of pneumoconiosis.It is expected to assist imaging doctors to diagnose pneumoconiosis by stages,reduce the rate of missed diagnosis and misdiagnosis,and improve the accuracy. |
PDF Full Text Request