Global incidence rate of non-tuberculous mycobacteria (NTM) pulmonary disease has been increasing rapidly (1,2). In the United States, the infection rate and prevalence of NTM exceed those of tuberculosis (TB) (3-5). NTM pulmonary disease is globally prevalent. NTM is currently a key public health problem (6). The growth of NTM is slower than that of pulmonary TB, making the disease course longer. Further, the overall cure rate is low due to resistance to first-line anti-TB drugs. NTM predominantly affects the lungs (7), with patients exhibiting chronic basic diseases and low immunity being more susceptible to NTM pulmonary disease. The imaging manifestations and clinical features of NTM are similar to TB, resulting in a high rate of misdiagnosis (7). If patients with NTM pulmonary disease are treated for TB, the treatment opportunity will be delayed (8). Sputum mycobacterium culture and strain identification are currently the primary methods employed to distinguish NTM pulmonary disease from TB. However, both methods are time-consuming, arduous and require strict laboratory standards (9). Additionally, the identification of strains is dependent upon a positive sputum mycobacterium culture, which poses a diagnostic challenge. However, the employment of computed tomography (CT) offers advantages, including rapid imaging, widespread availability and high-resolution density, thereby enabling feasible discrimination between diseases (10). Despite the advantages of CT, due to the high similarity of the pathogenesis and pathological features of NTM and TB, there is no reliable image feature to distinguish them (11). Consolidation is a common and important CT feature of NTM and TB (12). Consolidation characteristics of the diseases are similar, making it difficult to distinguish with the naked eye. CT-based radiomics has exhibited potential in the diagnosis and differentiation of pulmonary diseases through high-throughput extraction and mining data features (13,14). Due to differences in pathology between these two diseases (15), the possibility of identifying the diseases is further enhanced. In the present study, the value of CT-based radiomics analysis of consolidation characteristics in differentiation of NTM pulmonary disease from TB was investigated. The present study aimed to establish an effective classifier to provide a novel simple diagnosis method for NTM and TB.

A total of 381 ROIs (203 NTM pulmonary disease and 178 TB) were manually outlined (Fig. 1A and B) in the CT images of 192 patients (Table I). Computer-generated random numbers were utilized to assign 70% of ROI to the training data cohort and 30% to the verification data cohort. Using the variance threshold method, 456 features were selected for the models. Subsequently, best K method was employed in the selection of 315 features and then the LASSO algorithm was used to select 52 optimal features (Fig. 2). Of 52 radiomics features, 34 were texture analysis, one was shapes and 17 were first-order statistical feature groups. ROC curve analysis for both the training and validation cohort for differentiating between NTM pulmonary disease and TB is illustrated in Figs. 3 and 4.

Three classifiers were used to analyze the characteristics of the radiomic AUC; 95% CI, sensitivity and specificity of the training cohort and verification cohort are shown in Table II. In the training cohort, AUC [95% confidence interval (CI)], sensitivity and specificity of the KNN, SVM, and LR classifiers were 0.98 (0.95-1.00), 0.93 and 0.94, 0.99 (0.97-1.00), 0.96 and 0.96 and 0.99 (0.97-1.00), 0.97 and 0.97, respectively. In the validation cohort, these were 0.90 (0.82-0.97), 0.88 and 0.83, 0.92 (0.84-1.00), 0.86 and 0.85 and 0.89 (0.81-0.96), 0.88 and 0.85, respectively. In the external verification cohort, these were 0.84 (0.66-0.82), 0.65 and 0.83, 0.90 (0.78-0.92), 0.94 and 0.77 and 0.95 (0.84-0.96), 0.94 and 0.87, respectively. The ROC curves of the three classifiers are shown in Figs. 3 and 4. All cohorts had a significant AUC value >0.84.

In the training cohort, the precision of the three models was >0.94 and recall rate and F1 score were >0.93 (Table III). In the validation cohort, the precision of the three models was >0.83 with a recall rate and F1 score >0.86. Combining the precision, recall and F1 scores revealed that LR outperformed the other classifiers. In the external verification cohort, the precision of the three models was >0.77, the recall rate was >0.65 and F1 score was >0.69. LR classifier had the highest precision, recall and F1 score.

NTM is a conditional pathogen caused by Mycobacteria other than M. tuberculosis complex and Mycobacterium leprae, which commonly exist in water and soil. A total of 191 species has been discovered, but only a few can cause disease (17,26-28). Individuals typically contract the disease via the environment, with water and soil being key (7,26,27). Additionally, the incidence rate of NTM pulmonary disease is increasing in United States, Germany, Canada and Taiwan (29-32). The clinical symptoms and pathology of NTM pulmonary disease are difficult to distinguish from TB and NTM is prone to natural drug resistance. The diagnosis of NTM pulmonary disease and TB is achieved via etiological detection. However, this is slow and complex, negatively impacting clinical treatment (17). In the misdiagnosis of NTM pulmonary disease as TB, the use of anti-TB treatment results in the delay of proper treatment, a prolonged course of disease, poor prognosis and possible treatment failure (33). Therefore, discovering simple and effective diagnostic methods is necessary.

Conventional CT is one of the primary detection methods for NTM pulmonary disease and pulmonary TB. However, CT manifestations of NTM pulmonary disease are complex and typically perceived as conditions such as consolidation, bronchiectasis, cavities and bronchial dissemination, which are difficult to distinguish from TB (34,35). Koh et al (36) discovered that bronchiolitis, consolidation and bronchiectasis can be perceived on CT images of NTM pulmonary disease, which often involve >5 lobes. If bronchiectasis involving the middle lobe of the right lung and the tongue segment of the upper lobe of the left lung is observed in CT images, along with cavities and nodules, NTM pulmonary disease should be considered (37). In some studies (11,15), probability of NTM consolidation is significantly lower than that of pulmonary TB. However, in other studies (12,34), the consolidation of conventional CT does not significantly differentiate NTM and TB.

Necrotizing granulomatous inflammation is the key characteristic lesion of tuberculosis, NTM pulmonary disease, Coccidiosis and cryptococcosis (37-41). NTM pulmonary disease is more prone to suppurative necrosis than TB, while pink and basophilic necrosis caused by TB are more common than NTM pulmonary disease (42). NTM pulmonary disease possesses more giant or bizarre multinucleated giant cells compared with TB (43). The aggregation of epithelial-like cells leads to proliferative granuloma in NTM and caseous necrosis is less than that in TB (15). NTM pulmonary disease may result in atypical lesions, in which tissue cell aggregation with no granuloma is seen, which is common in immunodeficient patients (44,45). Consolidation of NTM pulmonary disease is less common than TB, with low incidence, less granuloma and more suppurative necrosis. It is difficult to observe differences in the pathological features of NTM and TB in conventional CT images, resulting in decreased differential diagnosis of TB.

Radiomics has developed substantially in recent years. Radiomics involves converting medical images into high-dimensional images, extracting data features through quantitative high-throughput and the analysis of the data for decision support (46). The sensitivity and predictive value of radiomics is significant in screening of small pulmonary nodules and diagnosis, treatment and prognosis of lung cancer (47-49). Radiomics may also detect common inflammatory lesions (50,51). Deep learning has been employed (25,52,53) to distinguish NTM pulmonary disease and TB, along with use of the cavity characteristics of radiomics to distinguish the two diseases. Consolidation features are common CT features of NTM pulmonary disease and TB. CT data extracted via high-throughput radiology reflect differences in the pathological characteristics of NTM and TB, compensating for fewer differences observed by the naked eye and the loss of information.

Here, consolidations of NTM pulmonary disease and TB were noted as ROIs. High-throughput image features were computed. A total of 52 radiomics features were obtained from the ROIs, of which 34 were texture analysis, one was shape and 17 were first-order statistical feature groups. Three supervised learning classifiers (KNN, SVM and LR) were used to analyze the extracted lung consolidation features. In the training cohort, the AUC values of models were all >0.98, 95% CI was 0.95-1.00, the sensitivity was >0.93 and the specificity was >0.94. In the validation cohort, the AUC values were all >0.89, 95% CI was 0.81-0.96, the sensitivity was >0.86 and the specificity was >0.83. AUC values of the ROC curve were all significant and their sensitivity and specificity as >0.83. In the external validation set, the AUC value was similar to that of internal validation set. In the present study, the characteristics of the classifiers were analyzed using clinical indicators (accuracy, recall, F1 score and support). The precision of models was >0.77, the recall rate was >0.65 and F1 score was >0.69. Further, the LR classifier yielded the highest precision, recall, and F1 score, which were >0.86, 0.87 and 0.87, respectively. From the results, it was observed that the radiomics features derived from consolidations held potential in differentiating between NTM pulmonary disease and TB. Although in some studies, the CT imaging characteristics of NTM pulmonary disease consolidations observed through traditional clinical methods differ from those of TB (16,40), results obtained by naked eye may be subjective. However, the present radiomics characteristics of consolidations have potential to distinguish between NTM pulmonary disease and TB. Radiomics analysis of consolidation characteristics of pulmonary diseases has the advantages of objectivity, quantification, stability and non-empirical dependence. Consequently, radiomics analysis possesses value in clinical application. Using radiomics characteristics to distinguish NTM pulmonary disease from TB is a promising, non-invasive and simple method. The early diagnosis of NTM pulmonary disease may improve the quality of life of patients and treatment of the disease, especially for resource-deficient medical systems in developing countries (54,55).

There are certain limitations to the present study. First, to ensure the homogeneity of the image, 5-mm-thick images were used, which may result in loss of information. Second, the sample size was small. Multicenter studies with larger sample sizes are required to validate the present results. ROI segmentation was performed manually, which may have been affected by subjective bias. Lastly, only consolidation was investigated and other characteristics were ignored, which may have resulted in incomplete information.

In the present study, radiomics features based on CT imaging were effective in identifying NTM pulmonary disease and TB consolidation. Additionally, LR classifier outperformed the other classifiers in the recognition of consolidation of NTM pulmonary disease in patients.