| BACKGROUNDIn recent years,the incidence rate of nontuberculous mycobacteria(NTM)pulmonary disease is increasing rapidly in the world.NTM grows slowly,has a long course of disease,and has a high resistance rate to first-line anti tuberculosis drugs,so the overall cure rate is low.NTM usually involves the lung.Its imaging and clinical features are similar to those of pulmonary tuberculosis(TB),and the misdiagnosis rate is high.At present,Mycobacterium sputum culture and identification are the main methods to identify NTM pulmonary disease and pulmonary tuberculosis,but these technologies are time-consuming and laborious,and require high laboratory level.Only when Mycobacterium sputum culture is positive,can we further identify the strains,which makes the diagnosis more difficult.Chest imaging examination is of great value in the diagnosis and differential diagnosis of NTM pulmonary disease and pulmonary tuberculosis.Because CT has the advantages of fast examination speed,universal use and high density resolution,it provides the feasibility for the identification of the two.However,due to the high similarity of the pathogenesis and pathological characteristics between the two,there is no reliable image feature to distinguish them.The CT manifestations of NTM pulmonary disease are cavity,bronchiectasis,bronchial dissemination and so on.It is difficult to differentiate NTM pulmonary disease from pulmonary tuberculosis.The imaging features of pulmonary tuberculosis are also multiple and multiple signs exist at the same time Although there are some similarities between NTM pulmonary disease and secondary pulmonary tuberculosis,there are some differences.The incidence rate of bronchiectasis in NTM is higher than that in pulmonary tuberculosis.In addition,the incidence of consolidation of two diseases,whether the cavity is located in the adjacent pleural membrane,whether pleural effusion and pleural tuberculoma are all in a certain statistical difference can be used as a basis for differential diagnosis,but it is also facing controversy.The CT findings of NTM pulmonary disease and pulmonary tuberculosis cavity have great similarity,so it is difficult to distinguish by naked eyes;for consolidation,the CT manifestations of both are more similar,and it is more difficult to identify with naked eyes.In recent years,artificial intelligence technology with deep learning as the core has made a series of major breakthroughs and has been widely used in various fields.At present,it is gradually applied to medical imaging,including CT imaging diagnosis of pulmonary diseases.At present,radiomics is widely used in the diagnosis of pulmonary diseases.With the rapid development of artificial intelligence radiomicss,it can transform medical images into high-dimensional images,mine data features through quantitative high-throughput mining,and then conduct data analysis for decision-making support.It has the characteristics of objectivity and refinement,and can find features that are difficult to detect by naked eyes.At present,there have been in-depth studies in the diagnosis of pulmonary diseases.For example,radiomics can be used to predict the histological characteristics,gene expression and prognosis of non-small cell lung cancer In this study,radiomics was used to extract the texture features of cavities and consolidation in NTM pulmonary disease and pulmonary tuberculosis CT images,and to explore the value of this technique in the differential diagnosis of NTM pulmonary disease and pulmonary tuberculosis CT images,and to provide a new and simple diagnostic method for clinical treatmentOBJECTIVETo explore the value of CT radiomics of the cavities and consolidation in the differential diagnosis of NTM pulmonary disease and Pulmonary tuberculosisMETHODSMethods:from February 2013 to March 2018 in Shandong Chest Hospital and Shandong University Qilu Hospital,73 patients with cavity NTM pulmonary disease and 73 patients with consolidation NTM pulmonary disease were retrospectively analyzed.At the same time,the same number of pulmonary tuberculosis patients with similar CT characteristics in the same time period were selected as the control group by random software.All images were Philips 64 row and gemstone CT chest plain scan images,with slice thickness of 5 mm.The DICOM format images were uploaded to Huiying radiology cloud platform v2.0 for processing.Cavity features::An experienced attending physician used double-blind method to observe and outline the CT images.In case of doubt,a senior physician reviewed the CT images to determine the regions of interest ROI.A total of 289 regions of interest ROIs cavity features were delineated from 146 patients in the cavity group,including 164 ROIs in NTM pulmonary diseases and 125 ROIs in pulmonary tuberculosis.80%of cavities were allocated to training data set and oyher 20%to verification data set by using random number generated by computer.Finally,131 ROIs of NTM pulmonary disease and 100 ROIs of tuberculosis were allocated to the training set,and 33 ROIs of NTM pulmonary disease and 25 ROIs of tuberculosis were allocated to the verification set.Firstly 476 features were selected from 1409 features using variance threshold method,then with the K best method,333 features were selected.Finally,we selected 24 optimal features with LASSO algorithmConsolidation features:An experienced attending physician used double-blind method to observe and outline the CT images.In case of doubt,a senior physician reviewed the CT images to determine the regions of interest ROI.A total of 246 regions of interest ROIs Consolidation features were delineated from 146 patients in the cavity group,including 108 ROIs in NTM pulmonary diseases and 138 ROIs in pulmonary tuberculosis.80%of consolidations were allocated to training data set and 20%to verification data set by using random number generated by the computer.Finally,86 ROIs of NTM pulmonary disease and 110 ROIs of pulmonary tuberculosis were allocated to the training set,and 22 ROIs of NTM pulmonary disease and 28 ROIs of tuberculosis were allocated to the verification set.Firstly 452 features were selected from 1409 features using variance threshold method,then with the K best method,203 features were selected.Finally,we selected 23 optimal features with LASSO algorithm.RESULTSCavity features:24 best features are selected by variance threshold method,K best method and lasso algorithm.Three supervised learning models(KNN,SVM,DT)are used for analysis.When training with KNN model,the AUC of training set were 0.99 in NTM(95%CI:0.96-1.00;sensitivity 0.94 and specificity 0.95),0.99 in TB(95%CI:0.96-1.00;sensitivity 0.95 and specificity 0.94),respectively,the AUC of validation set were 0.97 in NTM(95%CI:0.89-1.00;sensitivity 0.94 and specificity 0.84),0.97 in TB(95%CI:0.89-1.00;sensitivity 0.84 and specificity 0.94),respectively.When training with SVM model,the AUC of training set were 0.98 in NTM(95%CI:0.95-1.00;sensitivity 0.95 and specificity 0.95),0.98 in TB(95%CI:0.95-1.00;sensitivity 0.95 and specificity 0.95),respectively,the AUC of validation set were 0.99 in NTM(95%CI:0.90-1.00;sensitivity 0.94 and specificity 0.84),0.99 in TB(95%CI:0.90-1.00;sensitivity 0.84 and specificity 0.94),respectively.When training with DT model,the AUC of training set were 1.00 in NTM(95%CI:1.00-1.00;sensitivity 1.00 and specificity 1.00),1.00 in TB(95%CI:1.00-1.00;sensitivity 1.00 and specificity 1.00),respectively,the AUC of validation set were 0.85 in NTM(95%CI:0.76-0.95;sensitivity 0.91 and specificity 0.80),0.85 in TB(95%CI:0.76-0.95;sensitivity 0.80 and specificity 0.91),respectively.The AUC value of KNN model validation set is 0.97,that of SVM model is 0.99,and that of DT model is 0.85.The sensitivity and specificity of DT model verification set are also low.Through the analysis of accuracy,recall rate,F1 score and support,KNN and SVM models have better performance than DT model.Consolidation features:23 best features are selected by variance threshold method,K best method and lasso algorithm.Three supervised learning models(KNN,SVM,DT)are used for analysis.When training with KNN model,the AUC of training set were 0.95 in NTM(95%CI:0.88-1.00;sensitivity 0.90 and specificity 0.80),0.95 in TB(95%CI:0.88-1.00;sensitivity 0.80 and specificity 0.90),respectively,the AUC of validation set were 0.96 in NTM(95%CI:0.84-1.00;sensitivity 0.93 and specificity 0.80),0.96 in TB(95%CI:0.84-1.00;sensitivity 0.80 and specificity 0.93),respectively.When training with SVM model,the AUC of training set were 0.98 in NTM(95%CI:0.91-1.00;sensitivity 0.86 and specificity 0.88),0.98 in TB(95%CI:0.91-1.00;sensitivity 0.88 and specificity 0.86), respectively,the AUC of validation set were 0.97 in NTM(95%CI:0.86-1.00;sensitivity 0.93 and specificity 0.87),0.97 in TB(95%CI:0.86-1.00;sensitivity 0.87 and specificity 0.93),respectively.When training with DT model,the AUC of training set were 1.00 in NTM(95%CI:1.00-1.00;sensitivity 1.00 and specificity 1.00),1.00 in TB(95%CI:1.00-1.00;sensitivity 1.00 and specificity 1.00),respectively,the AUC of validation set were 0.77 in NTM(95%CI:0.61-0.93;sensitivity 0.80 and specificity 0.73),0.77 in TB(95%CI:0.61-0.93;sensitivity 0.73 and specificity 0.80),respectively.The AUC value of KNN model validation set is 0.96,SVM model validation set AUC value is 0.97,DT model validation set AUC value is 0.77.The sensitivity and specificity of DT model verification set are also low.The performance of KNN and SVM model through accuracy,recall rate,F1 score and support analysis is good,while DT model is not.CONCLUSIONCT radiomics to extract valuable cavity and consolidation features can make up for the lack of naked eye observation,which is of great significance in the differential diagnosis between nontuberculous mycobacterial pulmonary disease and pulmonary tuberculosis.KNN model and SVM model are more valuable than DT model in extracting cavity features and extracting consolidation features to distinguish NTM pulmonary disease from pulmonary tuberculosis.In the differential diagnosis of NTM pulmonary disease and pulmonary tuberculosis by radioomics,the feature extracted from the cavity is more practical and valuable. |
PDF Full Text Request