Whole-lung computed tomography radiomics combined with clinical features for differentiating multidrug-resistant tuberculosis from drug-sensitive tuberculosis: a retrospective multi-center study

Introduction

Pulmonary tuberculosis (PTB) is the leading cause of death from a single infectious disease and is exacerbated by the increasing prevalence of drug-resistant tuberculosis (DR-TB) (1). Multidrug-resistant tuberculosis (MDR-TB), a specific subtype of DR-TB, is characterized by resistance to at least rifampicin and isoniazid (2). Phenotypic drug susceptibility testing (pDST) is regarded as the gold standard for diagnosing DR-TB, however, it is time-consuming and may yield unreliable results owing to microbial undergrowth or other technical constraints (3). In July 2023, the World Health Organization (WHO) underscored the importance of targeted next-generation sequencing (tNGS) for detecting drug resistance in Mycobacterium tuberculosis (M. tuberculosis) (4). However, its high laboratory costs and technical demands have impeded its widespread application in clinical settings, thereby limiting its effectiveness in diagnosing and treating MDR-TB (5). Consequently, there remains a critical need for more efficient and accessible methods for early detection and management of DR-TB to improve treatment outcomes and curb transmission.

Chest computed tomography (CT) is the primary imaging modality for diagnosing DR-TB and monitoring treatment response (2). Compared with patients with drug-sensitive tuberculosis (DS-TB), those with MDR-TB typically exhibit a more extensive lesion distribution, with certain radiological features being more pronounced. For instance, cavitation is more frequently observed in MDR-TB and frequently manifests as thick-walled cavities (6). However, despite certain characteristic radiological findings associated with MDR-TB, the overall imaging presentation on chest CT scans remains largely similar between patients with MDR-TB and DS-TB patients, making differentiation challenging. Radiomics has introduced a novel approach to leverage imaging data for PTB diagnosis and classification, offering potential improvements in diagnostic accuracy and disease characterization. Radiomics has demonstrated potential for enhancing TB diagnosis (7,8).

Li et al. recently demonstrated that radiomics signatures extracted from focal cavity, tree-bud and nodular lesions can aid MDR-TB detection (9,10); however, these studies were limited to predefined local regions of interest (ROI). PTB is a chronic pathological condition characterized by multiple lesion morphologies and the coexistence of lesions at various stages of development. This is particularly evident in MDR-TB, where the lesions frequently exhibit a more extensive and even diffuse distribution across the lungs (2). A previous study has reported that lung lobe morphology has been used as an ROI for the radiomic analysis in certain diffuse lung diseases, including interstitial lung disease, offering novel insights into the study of pulmonary diseases, particularly those with diffuse patterns. Shen et al. (11) conducted a single-center retrospective study using whole-lung radiomics and a random forest model to differentiate non-tuberculous mycobacteria lung disease (NTM-LD) from PTB, achieving an area under the curve (AUC) of 0.821 in the validation cohort. This investigation provides a valuable reference for the application of radiomics in PTB.

To date, no study has investigated the application of whole-lung radiomics for diagnosing MDR-TB using chest CT scans. Consequently, this study aimed to assess the performance of radiomics features based on whole-lung ROI in distinguishing MDR-TB from DS-TB and to explore the added value of integrating clinical features for differential diagnosis. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1405/rc).

Methods

Study design and population

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethical Review Committee of The Fourth People’s Hospital of Nanning, Nanning, Guangxi, China {No. [2022] 64} and the Ethics Committee of The First People’s Hospital of Qinzhou, Qinzhou, Guangxi, China (No. KY20240110). As a retrospective study, the requirement for informed consent was waived and all data were de-identified. Hospital 1 is a Grade-A tertiary referral center for infectious diseases, and Hospital 2 is a Grade-A tertiary general hospital with a mixed-case population. Inpatients with culture-confirmed PTB and complete follow-up were screened. In the present study, the clinical dataset comprised only those variables that were routinely recorded during standard diagnostic and therapeutic workflows. Consequently, a complete-case approach was adopted for the missing data. MDR-TB was defined as in vitro resistance to both rifampicin and isoniazid by using pDST or tNGS assays, whereas DS-TB was defined as phenotypic or genotypic susceptibility to rifampicin, isoniazid, pyrazinamide, ethambutol and streptomycin. The inclusion criteria were as follows: (I) patients with a sputum culture-confirmed diagnosis of M. tuberculosis infection; (II) patients for whom a definitive differentiation between DS-TB and MDR-TB could be made based on available pDST or tNGS results; (III) patients with complete clinical and imaging data; and (IV) patients who completed follow-up registration. The exclusion criteria included the following: (I) patients with incomplete medical records or incomplete follow-up registration; (II) patients with human immunodeficiency virus (HIV)/acquired immunodeficiency syndrome (AIDS); (III) patients with a history of malignancy; (IV) patients with co-morbid conditions, including pneumoconiosis, pulmonary alveolar proteinosis, or non-tuberculous mycobacterial (NTM) infections; and (V) patients with poor image quality that compromised diagnostic accuracy. Due to the limited sensitivity of M. tuberculosis culture and the possibility of acquired resistance emerging during treatment, we excluded patients whose initial pDST was negative but subsequently developed mutations conferring resistance to any first-line medication. Two TB physicians and one thoracic radiologist reviewed the complete clinical records of all prospective cases, and an agreement among the three reviewers established the final eligibility for study inclusion.

Between November 2021 and June 2023, 627 eligible patients were enrolled in Hospital 1, including 195 with MDR-TB and 432 with DS-TB. These were randomly divided into the training set and the internal validation sets at a ratio of 8:2. The complete Hospital 2 cohort, comprising 123 patients (18 with MDR-TB and 105 with DS-TB) admitted between January 2021 and December 2023, was exclusively reserved for external validation. The comprehensive patient recruitment flow is illustrated in Figure 1.

Figure 1 The patient-recruitment flow of the study. Hospital 1: a Grade-A tertiary referral center for infectious diseases. Hospital 2: a Grade-A tertiary general hospital. AIDS, acquired immunodeficiency syndrome; DS-TB, drug-susceptible tuberculosis; HIV, human immunodeficiency virus; M. tuberculosis, Mycobacterium tuberculosis; MDR-TB, multidrug-resistant tuberculosis; NTM, nontuberculous mycobacteria; PTB, pulmonary tuberculosis; pDST, phenotypic drug susceptibility testing; tNGS, targeted next-generation sequencing.

All participants underwent non-contrast CT using a variety of scanners, including the Revolution CT ES, Optima CT680 (GE Healthcare, Milwaukee, WI, USA), SOMATOM Definition Flash, SOMATOM go.Top (Siemens Healthcare, Erlangen, Germany), and uCT760 (United Imaging Healthcare, Shanghai, China). The detailed acquisition parameters, including matrix, collimation, pitch, slice thickness, field of view, tube voltage, and tube current, for each scanner are provided in Table S1. Imaging was performed from the thoracic inlet to the bilateral adrenal glands using the deep inspiration breath-hold technique. Axial CT images of the entire thorax were obtained at full inspiration.

Collection of clinical data and the clinical model development

Clinical data were extracted from the hospital information system by one radiologist who was blinded to the final group assignment. The collected information included age, sex, history of tuberculosis (TB) treatment, diabetes mellitus (DM), body mass index (BMI), white blood cell (WBC) count, lymphocyte (LYMPH) count, monocyte (MONO) count, and neutrophil (NEUT) count. Routine blood tests must be conducted within 3 days before or after the first CT examination. Statistically significant variables were identified using univariate and multivariate logistic regression analysis. These variables were then used to develop a clinical model using the LightGBM algorithm, with n_estimators =2 and max_depth =2.

Image data acquisition and pre-processing

All images in the DICOM format were exported from the the hospital information system and subsequently converted to the nii.gz format, followed by a standardization process. This process included resampling the images to a voxel size of 1 mm × 1 mm × 1 mm and adjusting the gray values to a window level of −500 Hounsfield unit (HU) and a window width of 1,500 HU. These adjustments were made to mitigate the effects of varying the layer thicknesses and to reduce noise interference. A publicly accessible U-net model (LTRCLobes_R231; https://github.com/JoHof/lungmask), a deep learning model designed for automated lung segmentation in CT scans with severe pathologies, was then used to segment the right and left lungs automatically (12). A thoracic radiologist with 13 years of experience used the ITK-SNAP software (version 3.8.0; www.itksnap.org) to rectify any erroneous segmentation and conducted a second recalibration one month after the initial correction. The corrected segmented images were then fused with the whole-lung images using the SimpleITK package (version 2.4.0) in Python.

Whole lung radiomics feature extraction

Radiomics features were extracted from the original CT images and from eight wavelet-filtered decompositions, three Laplacian-of-Gaussian (LoG) scales (σ =1.0, 2.0 and 3.0 mm), as well as square, squareRoot, logarithm, exponential, and gradient filtered images using PyRadiomics (version 3.0.1). For each image type, the following feature classes were extracted: shape, first-order statistics, neighbouring gray tone difference matrix (NGTDM), gray-level size zone matrix (GLSZM), gray-level run-length matrix (GLRLM), gray-level dependence matrix (GLDM), and gray-level co-occurrence matrix (GLCM). All features were subsequently standardized using Z-score normalization before further analysis.

Radiomics feature dimension reduction, selection and model construction

Optimal radiomics features were selected using a three-step process. First, the normality of each continuous radiomics feature was first assessed using the Shapiro-Wilk test. Features that were normally distributed were compared between groups using an independent-samples t-test; non-normally distributed features were compared using the Mann-Whitney U test. Features with P≥0.05 were considered statistically insignificant and were excluded from further analysis. The Pearson correlation coefficient was then used to assess the inter-feature correlation. Features exhibiting a correlation coefficient ≥0.9 were examined, and the feature with the highest average absolute correlation was excluded from further analysis. Finally, the least absolute shrinkage and selection operator (LASSO) algorithm was applied to the training cohort data to eliminate redundant and irrelevant parameters. The penalty parameter was optimized using ten-fold cross-validation, and features with non-zero coefficients were retained for the final radiomics prediction model. To comply with the events-per-variable principle (≥10), the number of candidate predictors was capped at one-tenth of the training-set positive events. The radiomics model was then constructed using the multilayer perceptron (MLP) algorithm. For MLP we set hidden_layer_sizes =128, 64, 32, max_iter =10, solver = ‘adam’, and random_state =123. These configurations aimed to optimize the training efficiency and prediction performance, ensuring rapid model convergence and the effective handling of linearly separable data.

The training cohort was used for model development, whereas the internal and external validation cohorts were preserved as independent entities for evaluating the classification performance of the developed models. After training, the optimal model was selected and evaluated using the test set, to estimate the probability of MDR-TB, ranging from 0% to 100%. The clinical and radiomics models were integrated to form a combined model. A nomogram was then generated to visualize the combined model, graphically assess the significance of the variables, and calculate its identification accuracy. Crucial evaluation metrics included the AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. Receiver operating characteristic (ROC) curves were plotted to visually represent model performance. The entire model-building process was implemented on the Python-based OnekeyAI platform (https://github.com/OnekeyAI-Platform/onekey). Figure 2 illustrates the flowchart of the study.

Figure 2 The work flow of the study. BMI, body mass index; CT, computed tomography; DM, diabetes mellitus; GLCM, gray-level co-occurrence matrix; GLDM, gray-level dependence matrix; GLRLM, gray-level run-length matrix; GLSZM, gray-level size zone matrix; HU, Hounsfield unit; LYMPH, lymphocyte; MONO, monocyte; NGTDM, neighbouring gray tone difference matrix; NEUT, neutrophil; WBC, white blood cell; WL, window level; WW, window width.

Statistical analysis

Statistical analysis was conducted using the Statistical Package for the Social Sciences software (version 27.0). The normality of the clinical variables was assessed using the Shapiro-Wilk test. Continuous variables were analyzed using the independent samples t-test or the Mann-Whitney U test, while categorical variables were analyzed using the Chi-squared test. A significance level of P<0.05 was considered statistically significant.

Results

Clinical feature screening and clinical model development

This study analyzed a total of 750 patients, including 213 patients with MDR-TB and 537 patients with DS-TB. The clinical characteristics of the patients in training and validation cohorts are presented in Table 1. No statistically significant differences were observed among the groups with respect to age, gender, BMI, DM, WBC count, LYMPH count, MONO count, or NEUT count. The incidence of MDR-TB was higher in the relapsed patients than in the newly treated patients across all cohorts. The differences were statistically significant. The independent risk factors identified through univariate and multivariate logistic regression analyses are detailed in Table 2. Age (P<0.001), sex (P<0.001), BMI (P<0.001), DM (P=0.001), TB retreatment (P=0.001), and WBC count (P<0.001), LYMPH count (P<0.001), and MONO count (P<0.001) were identified as the potential risk factors. Multivariate regression analysis further identified DM [P=0.02; odds ratio (OR) =1.901; 95% confidence interval (CI): 1.223, 2.954] and TB retreatment (P<0.001; OR =11.51; 95% CI: 7.710, 17.725) as independent risk factors for MDR-TB.

A clinical model was constructed using the LightGBM algorithm, demonstrating moderate discriminatory performance with AUC values of 0.742, 0.738, and 0.725 in training, internal validation, and external validation sets, respectively.

A radiomics model and a nomogram model development

A comprehensive feature-extraction process yielded 1,832 features. Following univariate filtering with the t-test, Mann-Whitney U-test, and Pearson correlation analysis, 322 features remained. LASSO was then applied to identify an optimal subset for final model construction (Figure 3A,3B). A radiomic signature was established using seven features (Figure 3C). The radiomics model was constructed using the MLP algorithm, achieving AUC values of 0.724, 0.720, and 0.703 for training, internal validation, and external validation sets, respectively.

Figure 3 Radiomics feature selection workflow. (A) Ten-fold cross-validation used to select the optimal λ in the LASSO regression; the vertical dashed line marks λ =0.03913, which yielded the smallest mean-squared error. (B) Coefficient path showing the shrinkage trajectories of all radiomics features; the seven variables that retained non-zero coefficients are highlighted. (C) Bar plot of the seven retained radiomics features and their corresponding LASSO coefficients. λ: penalty tuning parameter. LASSO, least absolute shrinkage and selection operator; MSE, mean squared error.

To further improve discriminative performance, we integrated the clinical model with the radiomics model into a combined model. The combined model achieved AUCs of 0.816, 0.795, and 0.835 in the training, internal-validation and external-validation sets, respectively (Table 3 and Figure 4A-4C). DeLong tests verified that the combined model significantly outperformed both constituent models in the training cohort (both P<0.001). In the internal-validation cohort it surpassed the radiomics model alone (P=0.04), whereas in the external-validation cohort it exceeded the clinical model (P=0.03). A user-friendly nomogram for clinical application is presented in Figure 5.

Figure 4 The ROC curves of the radiomics model, clinical model, and combined model for predicting MDR-TB from DS-TB patients. (A) The training cohort. (B) The internal validation cohort. (C) The external validation cohort. AUC, area under the curve; CI, confidence interval; DS-TB, drug-susceptible tuberculosis; MDR-TB, multidrug-resistant tuberculosis; ROC, receiver operating characteristic.

Figure 5 The nomogram for prediction of MDR-TB based on clinical characteristics and Radiomic score. Instructions: the nomogram comprises three variable scales: DM (absent =0, present =1), TB retreatment: (treatment-naive =0, retreatment =1), and a seven-feature radiomics risk score (range, 0–0.9; linearly converted to 0–100 points). Summing the points from all three axes yields the total score, which is then projected downward to the bottom scale to obtain the individualized probability of the outcome. DM, diabetes mellitus; MDR-TB, multidrug-resistant tuberculosis; TB, tuberculosis.

Discussion

This retrospective study developed a combined model using whole-lung CT radiomics features and clinical characteristics to differentiate between DS-TB and MDR-TB. The combined model, which integrated whole-lung radiomics and clinical features, exhibited superior classification performance, with AUC values of 0.816, 0.795, and 0.835 in training, internal validation, and external validation sets, respectively. These findings underscore the potential of whole-lung radiomics to improve the diagnosis of MDR-TB.

In this study, no statistically significant difference in age was observed between the patients with MDR-TB and those with DS-TB. This finding contradicts prior study which suggested that patients with MDR-TB are generally younger (9). This discrepancy may be attributed to a patient selection bias. Similarly, no statistically significant differences in gender or BMI were observed between the two groups, consistent with prior study (9). Previous studies have indicated that DR-TB is more prevalent among patients who have undergone retreatment, experienced intermittent treatment, or exhibited disease progression (13,14). Our findings corroborated these observations. The WHO report indicates that in 2023, the estimated prevalence of MDR-TB/rifampicin-resistant tuberculosis (RR-TB) cases among new TB cases in 2023 was 3.2% [95% uncertainty interval (UI): 2.5–3.8%], while among previously treated cases, it was 16% (95% UI: 9.0–24.0%) (1). In our study, the proportion of MDR-TB cases in both the training and internal validation cohorts was higher than the global average. This discrepancy is primarily attributed to selection bias in case enrollment. Our study included patients with DS-TB who had completed follow-up registration, which inadvertently increased the proportion of patients with MDR-TB. Moreover, The Fourth People’s Hospital of Nanning serves as a designated treatment center for MDR-TB, leading to a higher-than-average proportion of MDR-TB cases compared with real-world settings. In the external validation cohort, the prevalence of MDR-TB among previously treated patients was 35.29% (12/34), whereas among newly treated patients, it was 6.74% (6/89), approximately twice the global average. This elevated incidence may be due to the relatively high prevalence of MDR-TB in our country. This may also be associated with the exclusion of patients with pure RR-TB in our study population.

Our study revealed no significant difference in the prevalence of DM between patients with MDR-TB and DS-TB patients, consistent with previous findings (14). However, certain studies have revealed that the coexistence of type 2 DM may enhance the transmissibility and activity of PTB, as well as elevating the prevalence of drug resistance (15,16). This discrepancy could be attributed to differences in the geographical areas of the study populations. Furthermore, a prior study has indicated that WBC, MONO, and NEUT counts are elevated, while LYMPH count is significantly reduced in individuals with active TB compared with those with latent TB infection (17). In our study, no statistically significant differences in WBC, LYMPH, MONO, and NEUT counts were observed between patients with MDR-TB and DS-TB, potentially because of similar immune responses elicited in active TB.

Earlier studies on ROI selection demonstrated that models based on various ROIs exhibited varying diagnostic performances (18). Li et al. (19) and Cui et al. (20) discovered that incorporating perilesional information could improve classification performance in radiomics studies. Liu et al. (21) developed radiomics models to differentiate DR-TB from DS-TB using CT-derived features extracted separately from the pulmonary cavity, tree-in-bud sign, total lung lesions, and residual pulmonary parenchyma, and a composite features of all of the above. Their results indicated that the composite-feature model achieved marginally superior diagnostic performance compared with each single-region model. Previous studies have reported that lung lobe morphology can be used as an ROI for radiomic analysis in certain diffuse lung diseases, including interstitial lung disease, offering novel insights into the study of pulmonary diseases. Jiang et al. (22) developed a radiomics model using whole-lung radiomics features to predict the gender-age-physiology (GAP) staging in patients with connective tissue disease-associated interstitial lung disease (CTD-ILD). The model achieved AUC values of 0.803 and 0.801 in the training and validation cohorts, respectively, indicating a strong discriminatory power between GAP stages I and II/III. Other researchers have extracted radiomic features from the right lung region and integrated them with clinical and CT-based semantic features to construct a nomogram for GAP staging in CTD-ILD. The AUC values of this model in the training and validation cohorts reached 0.995 and 0.851, respectively (23). Lin et al. (24) developed a nomogram based on whole-lung radiomics to predict cardiovascular disease risk in patients with chronic obstructive pulmonary disease, achieving AUC values of 0.731, 0.727, and 0.725 in training, internal validation, and external validation cohorts, respectively.

Considering the extensive lung parenchymal damage and significant regional heterogeneity in MDR-TB (2), whole-lung assessment can more accurately reflect the overall disease characteristics. Moreover, we employed automatic segmentation followed by manual correction to ensure the reproducibility of ROI delineation. This semi-automated approach can yield comparable outcomes across observers with varying levels of expertise (25).

In this study, the whole-lung radiomics model selected seven radiomic features, Five first-order features, one GLSZM and one GLRLM feature. First-order features primarily utilize basic statistical metrics to describe the distribution of voxel intensities within the ROI, whereas GLRLM features measure gray-level runs, defined as the length (number of consecutive pixels) of pixels sharing the same gray-level value (26). Consistent with previous studies, our findings reveal heterogeneity in MDR-TB and DS-TB, with several textural features being associated with MDR-TB (9,10). These results indirectly associated with high-dimensional spatial information, potentially assisting in the differential diagnosis.

Compared to studies that focused on radiological semantic features, including cavitation (9), tree-in-bud patterns and nodules (10) as ROIs, the radiomics model in our study demonstrated relatively lower performance. However, radiomics studies that depend upon specific lesion-based ROIs frequently require substantial expertise from researchers and can exhibit variable reproducibility. Conversely, our study employed the entire lung as the ROI, utilizing an automatic lung segmentation technique based on the Lungmask. This approach has fewer requirements for researchers and is more accessible. Moreover, several mature tools are currently accessible for whole-lung or whole-organ segmentation, including TotalSegmentator, MONAI Auto3dseg, and Lungmask, which offer precise automatic segmentation and facilitate future research.

This study evaluated the efficacy of whole-lung CT radiomics, clinical, and combined models in diagnosing MDR-TB. Across all cohorts, the combined model consistently outperformed both the clinical and radiomics models, particularly regarding AUC, sensitivity, and specificity. These findings revealed that integrating clinical and radiomic features can significantly enhance the diagnostic accuracy of MDR-TB. Although the radiomics model demonstrated superior sensitivity compared to the clinical model, its lower specificity may result in an increased rate of false positives. Conversely, the clinical model, despite its ease of implementation and higher specificity, may overlook several MDR-TB cases due to its lower sensitivity. The balanced performance of the combined model in sensitivity and specificity, along with its stability across various validation cohorts, renders it a promising tool for clinical applications. Subsequent studies must focus on enhancing this model and exploring its implementation in real-world clinical settings to improve its diagnostic efficacy. This study’s findings underscore the potential of integrating clinical and radiomics models to improve the accuracy of MDR-TB diagnoses, which may significantly impact patient management and public health measures. Several scholars have highlighted the significance of evaluating both the diagnostic yield and the coverage of TB diagnoses (27). They proposed that moderately sensitive TB tests, which utilize more accessible specimens, may diagnose a larger number of individuals than highly sensitive molecular tests that depend on sputum samples. CT is an essential tool for TB diagnosis and treatment monitoring, offering extensive coverage at a comparatively low cost. Although it exhibits moderate sensitivity and specificity, our study demonstrates its potential for clinical application in diagnosing MDR-TB.

This study has several limitations. First, selection bias may be inherent; the derivation cohort was recruited from a specialized infectious-disease hospital, whereas the validation cohort was obtained from general hospitals, resulting in a significant imbalance in the proportion of MDR-TB and DS-TB cases. This discrepancy may compromise the model performance assessment. Second, the number of MDR-TB cases in internal and external validation sets fell short of the recommended 100-event threshold. Consequently, these results should be regarded as exploratory, and larger multi-center studies are required for definitive validation. Whole-lung ROIs were initially segmented using a deep-learning U-Net model (LTRCLobes_R231) and then manually refined. Given the frequent occurrence of parenchymal destruction, volume loss, and pleural effusion in patients with TB, this manual repair process is labour-intensive.

Conclusions

In summary, we developed models utilizing whole-lung CT radiomics, clinical characteristics, and their combinations to evaluate the risk of MDR-TB in patients with PTB. Our findings underscore the potential of whole-lung CT radiomics and clinical attributes for diagnosing MDR-TB. However, the generalizability of these models to independent external datasets necessitates additional optimization.

Acknowledgments

The authors extend thanks to all the clinicians who assisted in this study.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1405/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1405/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1405/prf

Funding: This work was supported by the Key Research and Development Program of Nanning (No. 20233069), and the Scientific Research and Technology Development Projects of Xingning District (No. 2022A11).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1405/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethical Review Committee of The Fourth People’s Hospital of Nanning, Nanning, Guangxi, China {No. [2022] 64} and the Ethics Committee of The First People’s Hospital of Qinzhou, Qinzhou, Guangxi, China (No. KY20240110). As a retrospective study, the requirement for informed consent was waived and all data were de-identified.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. WHO. Global tuberculosis report. Available online: https://www.who.int/teams/global-programme-on-tuberculosis-and-lung-health/tb-reports/global-tuberculosis-report-2024
2. Xu CJ, Lu PX, Li CH, et al. Chinese expert consensus on imaging diagnosis of drug-resistant pulmonary tuberculosis. Quant Imaging Med Surg 2024;14:1039-60. [Crossref] [PubMed]
3. Li S, Chen W, Feng M, et al. Drug Resistance and Molecular Characteristics of Mycobacterium tuberculosis: A Single Center Experience. J Pers Med 2022;12:2088. [Crossref] [PubMed]
4. WHO. Use of targeted next-generation sequencing to detect drug-resistant tuberculosis Rapid communication. Available online: https://iris.who.int/bitstream/handle/10665/371687/9789240076372-eng.pdf?sequence=1
5. de Araujo L, Cabibbe AM, Mhuulu L, et al. Implementation of targeted next-generation sequencing for the diagnosis of drug-resistant tuberculosis in low-resource settings: a programmatic model, challenges, and initial outcomes. Front Public Health 2023;11:1204064. [Crossref] [PubMed]
6. Lee H, Sohn JW, Sim YS, et al. Outcomes of extended duration therapy for drug-susceptible cavitary pulmonary tuberculosis. Ann Transl Med 2020;8:346. [Crossref] [PubMed]
7. Yang L, Jiang Z, Tong J, et al. Development and validation of a preoperative CT based radiomics nomogram to differentiate tuberculosis granulomas from lung adenocarcinomas: an external validation study. BMC Cancer 2024;24:670. [Crossref] [PubMed]
8. Yan Q, Zhao W, Kong H, et al. CT based radiomics analysis of consolidation characteristics in differentiating pulmonary disease of non tuberculous mycobacterium from pulmonary tuberculosis. Exp Ther Med 2024;27:112. [Crossref] [PubMed]
9. Li Y, Wang B, Wen L, et al. Machine learning and radiomics for the prediction of multidrug resistance in cavitary pulmonary tuberculosis: a multicentre study. Eur Radiol 2023;33:391-400. [Crossref] [PubMed]
10. Li Y, Xu Z, Lv X, et al. Radiomics analysis of lung CT for multidrug resistance prediction in active tuberculosis: a multicentre study. Eur Radiol 2023;33:6308-17. [Crossref] [PubMed]
11. Shen J, Yu M, Jin X, et al. Whole lung CT radiomics combined with clinical and conventional CT features for distinguishing non-tuberculous mycobacterial pulmonary disease and pulmonary tuberculosis. Chinese Journal of Interventional Imaging and Therapy 2025;22:16-21.
12. Hofmanninger J, Prayer F, Pan J, et al. Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem. Eur Radiol Exp 2020;4:50. [Crossref] [PubMed]
13. Cheng N, Wu S, Luo X, et al. A Comparative Study of Chest Computed Tomography Findings: 1030 Cases of Drug-Sensitive Tuberculosis versus 516 Cases of Drug-Resistant Tuberculosis. Infect Drug Resist 2021;14:1115-28. [Crossref] [PubMed]
14. Cai X, Xu X, He G, et al. Drug resistance in drug-resistant tuberculosis patients with and without diabetes mellitus: a comparative analysis. BMC Infect Dis 2024;24:807. [Crossref] [PubMed]
15. Guo J, Feng X, Pang J, et al. Risk factors for rifampicin-susceptible and isoniazid-resistant tuberculosis in adult patients with type 2 diabetes mellitus in Nanjing. BMC Infect Dis 2025;25:335. [Crossref] [PubMed]
16. Kumar V, Yadav J, Parmar A, et al. Study of rifampicin resistance among newly diagnosed pulmonary tuberculosis patients with type 2 diabetes mellitus: a prospective observational study. International Journal of Research in Medical Sciences 2021;9:2035-42.
17. Kang YJ, Park H, Park SB, et al. Combined analysis of whole blood interferon gamma release assay and complete blood count analysis for rapid discrimination of active tuberculosis and latent tuberculosis infection. J Clin Tuberc Other Mycobact Dis 2021;24:100253. [Crossref] [PubMed]
18. Dammak S, Gulstene S, Palma DA, et al. Distinguishing recurrence from radiation-induced lung injury at the time of RECIST progressive disease on post-SABR CT scans using radiomics. Sci Rep 2024;14:3758. [Crossref] [PubMed]
19. Li Y, Lyu B, Wang R, et al. Machine learning-based radiomics to distinguish pulmonary nodules between lung adenocarcinoma and tuberculosis. Thorac Cancer 2024;15:466-76. [Crossref] [PubMed]
20. Cui EN, Yu T, Shang SJ, et al. Radiomics model for distinguishing tuberculosis and lung cancer on computed tomography scans. World J Clin Cases 2020;8:5203-12. [Crossref] [PubMed]
21. Liu M, Zhou Y, Ding J, et al. Prediction of active drug-resistant pulmonary tuberculosis based on CT radiomics: construction and validation of independent models and combined models for residual pulmonary parenchyma. Front Med (Lausanne) 2025;12:1508736. [Crossref] [PubMed]
22. Jiang X, Su N, Quan S, et al. Computed Tomography Radiomics-based Prediction Model for Gender-Age-Physiology Staging of Connective Tissue Disease-associated Interstitial Lung Disease. Acad Radiol 2023;30:2598-605. [Crossref] [PubMed]
23. Qin S, Jiao B, Kang B, et al. Non-contrast computed tomography-based radiomics for staging of connective tissue disease-associated interstitial lung disease. Front Immunol 2023;14:1213008. [Crossref] [PubMed]
24. Lin X, Zhou T, Ni J, et al. CT-based whole lung radiomics nomogram: a tool for identifying the risk of cardiovascular disease in patients with chronic obstructive pulmonary disease. Eur Radiol 2024;34:4852-63. [Crossref] [PubMed]
25. Zhan Y, Wang Y, Zhang W, et al. Diagnostic Accuracy of the Artificial Intelligence Methods in Medical Imaging for Pulmonary Tuberculosis: A Systematic Review and Meta-Analysis. J Clin Med 2022;12:303. [Crossref] [PubMed]
26. Defeudis A, De Mattia C, Rizzetto F, et al. Standardization of CT radiomics features for multi-center analysis: impact of software settings and parameters. Phys Med Biol 2020;65:195012. [Crossref] [PubMed]
27. Broger T, Marx FM, Theron G, et al. Diagnostic yield as an important metric for the evaluation of novel tuberculosis tests: rationale and guidance for future research. Lancet Glob Health 2024;12:e1184-91. [Crossref] [PubMed]