Abstract: Objective To analyze the value of computer tomography(CT)radiomics features on differentiating nontuberculous mycobacteria(NTM)lung diseases with cavity from pulmonary tuberculosis with similar cavity. Methods Clinical data of 51 pulmonary NTM patients and 42 pulmonary tuberculosis patients with similar cavity from February 2013 to March 2018 in Shandong Provincial Chest Hospital and Qilu Hospital of Shandong University were retrospectively analyzed. Double-blind method was used to observe and sketch CT images, and 198 cavities of volume of interests(VOI)were drawn by two experienced radiologists, and then 80% of VOI cavities were allocated to training data set and 20% to verification data set by using random number generated by computer. A total of 1 409 radiomics features extracted from Radcloud platform were used to analyze the differences in CT cavity characteristics of the two diseases. The best features were selected by variance threshold method, K best method and Lasso algorithm. The receiver operating characteristic(ROC)curves were analyzed by three supervised learning classifiers(KNN, SVM and DT). Results A total of 94 best features were selected. The different learning classifiers showed that the lowest and the highest AUC values of validation set were 0.95 and 1.00, respectively. The sensitivity and specificity of the verification set were more than 0.95. The fine performance of the three classifiers was obtained by using four indicators(precision, recall rate, F1 score and support degree). Conclusion Some valuable cavity features can be extracted via CT radiomics, and are helpful for the differential diagnosis between the pulmonary NTM and pulmonary tuberculosis, which may make up for the lack of visual observation of the common CT images.

Key words: Radiomics, Cavity, Computer tomography, Nontuberculous mycobacteria lung diseases, Pulmonary tuberculosis