Study Of AI Aided Diagnosis Based On Manual Annotation CT Image Data Resources In Nontuberculous Mycobacteria Lung Disease

Part Ⅰ Distinguishing Nontuberculous Mycobacteria from Mycobacterium Tuberculosis Lung Disease from CT Images Using a Deep Learning FrameworkPurpose To develop and evaluate the effectiveness of a deep learning framework 3D-Res Net based on CT images to distinguish NTM-LD from MTB-LD.Method Chest CT images of 1105 patients with either 301 NTM-LD or 804 MTB-LD confirmed by pathogenic microbiology were retrospectively collected.3D-Res Net was developed by randomly extracting data in an 8:1:1 ratio for training,validating,and testing.Its performance was evaluated by accuracy,specificity,sensitivity,precision,and F1 score.The model was compared with three radiologists in the test set,and the clinical applicability was evaluated using confusion matrix,ROC,and AUC.Result The AUCs of our model on training,validating,and testing datasets were 0.903,0.876,and 0.856,respectively.Additionally,the performance of the model on AUC,accuracy,sensitivity,and specificity was higher than that of the radiologists.Meanwhile,without manual annotation,the model automatically identified lung diseases on chest CT,with an efficiency over 1000 times that of the radiologists.Conclusion This study shows the efficacy of the deep learning framework as a rapid auxiliary diagnostic tool for NTB-LD and MTB-LD.Its use can help provide timely and accurate treatment strategies to patients with these diseases.Part Ⅱ Comparative Chest Computed Tomography Findings of Nontuberculous Mycobacterial Lung Diseases and Mycobacterium Tuberculosis Lung DiseasePropose To compare and analyze the chest CT imaging characteristics and differences between patients with NTM-LD and MTB-LD.Methods Simple random sampling method was used to select 200 NTM-LD and 200 MTB-LD patients for retrospective analysis from January 2014 to January 2020 in Tianjin Haihe Hospital.Two radiologists read the CT imagines independently,evaluating and calculating the frequency of the CT signs in the chest images.Result The age was significantly higher(P<0.001)and wheezing symptom was more common in NTB-LD patients(P=0.017).Smoking,coughing and sputum symptoms were more common(P=0.001;P=0.007;P=0.009)and the incidence of diabetes was higher(P= 0.011)in MTB-LD patients.NTM-LD patients have a higher proportion of ground-glass opacity,bronchiectasis,thickened pleura,lymphadenopathy and calcification(P<0.001;P<0.001;P<0.001;P=0.012;P=0.043).The proportion of MTB-LD patients with consolidation,nodule,and cavity was higher(P=0.002;P=0.034;P=0.014),and the wall thickness of cavity in patients with NTM-LD was significantly smaller(P<0.001).Conclusion NTM-LD and MTB-LD patients are different in age,diabetes mellitus,clinical symptoms and CT features,which helps to distinguish NTM-LD from MTB-LD.The ratio and morphological features of bronchiectasis and cavities can be used to distinguish NTM-LD from MTB-LD,which can be used as a feature of artificial labeling machine learning.Part Ⅲ Machine Learning-Based Differentiation of NTM-LD and MTB-LD based on Manual Annotation CT ImagesPurpose To establish a machine learning algorithm based on CT that can identify non-tuberculous mycobacterial pulmonary disease and pulmonary tuberculosis.Method Using machine learning algorithms and manual annotations,57 cases of MTB-LD and 59 cases of NTM-LD were selected,and 103 quantitative features were extracted through machine learning algorithms.After comparing the parameters of the features,the parameters are selected,and the discriminative feature recognition is performed by linear support vector machine.Result The experimental results show that bronchiectasis has a greater impact on the prediction results,and due to its good prediction performance(AUC 0.84 ± 0.06;accuracy 0.85 ± 0.06;sensitivity 0.88 ± 0.07;specificity 0.80 ± 0.12),They can distinguish two diseases efficiently.Conclusion This exploratory study provides insight into machine learning based identification of NTM lung diseases.The algorithm of quantitative bronchiectasis feature data is the main factor that affects the classification,and two of the features(bronchiectasis and cavity)can be used to distinguish diseases.