关键词: 结核,肺, 分枝杆菌感染, 诊断,鉴别, 体层摄影术,X线计算机, 影像组学

Abstract:

Objective: To explore a machine learning model based on chest CT images for differential diagnosis of nontuberculous mycobacterium lung disease (NTM-LD) and pulmonary tuberculosis (PTB). Methods: Chest CT images of 120 patients (NTM-LD) and 120 patients (PTB) were retrospectively collected in Tianjin Haihe Hospital from January 2017 to December 2020. 168 cases (70%) were randomly selected as the training set, and 72 cases (30%) were selected as the testing set. Chest CT images of 25 patients (NTM-LD) and 25 patients (PTB) from Xi’an Chest Hospital were collected as an external validation set. A total of 12 radiologist semantic features and 2107 radiomic features were extracted from chest CT images, and 40 radiomic features were retained through feature dimensionality reduction. Three distinct machine learning classification models were constructed utilizing the Support Vector Machines (SVM) algorithm. These models encompass a semantic model, a radiomics model, and a hybrid radiomics-semantic model. The diagnostic performance of the three models were evaluated by the receiver operating characteristic (ROC) curve and the area under the curve (AUC). The statistical significance of differences between the three models were compared by DeLong test. Results: In the testing set, the AUC of radiomics-semantic model, radiomics model and semantic model were 0.9853, 0.9282, and 0.7901, respectively. There were statistically significant differences between semantic model and radiomics-semantic model, as well as between semantic model and radiomics model (Z=2.759, P=0.006; Z=2.230,P=0.026). However, there was no statistically significant difference between radiomics-semantic model and radiomics model (Z=0.761, P=0.502).In the external validation set, the AUC of radiomics-semantic model, radiomics model and semantic model were 0.9216, 0.9024 and 0.7624, respectively. There was a statistically significant difference between radiomics-semantic model and semantic model (Z=2.126,P=0.034). However, there was no statistically significant difference between radiomics-semantic model and radiomics model (Z=0.368,P=0.713). Conclusion: Compared with semantic model, the machine learning model combining radiomics and semantic features showed an excellent diagnostic efficiency and great clinical application value in distinguishing NTM-LD and PTB. Although its performance improvement was not significant compared to radiomics model.

Key words: Tuberculosis,pulmonary, Mycobacterium infections, Diagnosis,differential, Tomography, X-ray computed, Radiomics