Application of Clinical-Radiomics Fusion Model in Predicting Drug Resistance of Cavitary Pulmonary Tuberculosis

Introduction

Tuberculosis (TB) remains one of the deadliest infectious diseases worldwide, with more than 25% of the global population infected. In 2023, an estimated 10.8 million new cases and 1.25 million deaths were reported, making TB the leading cause of death from a single infectious agent.¹ Among the major challenges to TB control is drug-resistant TB (DR-TB), particularly multidrug-resistant TB (MDR-TB), defined as resistance to at least isoniazid and rifampin, the two key first-line anti-TB agents.² MDR-TB is characterized by prolonged treatment duration, significant adverse effects, high treatment costs, and lower cure rates than drug-sensitive TB (DS-TB), thereby impeding global efforts to end the TB epidemic.^3,4 The Global Burden of Disease Study has shown a continuous global rise in MDR-TB cases and related indicators from 1990 to 2021, with projections suggesting further escalation by 2050.⁵ Among the various clinical manifestations of TB, cavitary pulmonary TB holds a particularly critical position. The formation of lung cavities, often associated with high Mycobacterial tuberculosis (MTB) burden and enhanced transmissibility, indicates severe disease progression.^6,7 Cavities hinder the adequate penetration of anti-TB drugs, reducing treatment efficacy. Consequently, even with standard therapy, complete bacterial eradication may not be achieved, increasing the risk of drug resistance. Additionally, cavities disrupt local immune responses, bacterial clearance and predisposing patients to infection with drug-resistant strains or progression to MDR-TB.⁸ Therefore, early, rapid, and accurate prediction of drug resistance in cavitary pulmonary TB is crucial for optimizing treatment strategies, interrupting transmission, and improving clinical outcomes. Currently, the diagnosis of TB drug resistance relies on microbiological methods, including traditional phenotypic drug susceptibility testing (DST) and molecular biology-based techniques. Although phenotypic DST remains the “gold standard”, it is time-consuming, often requiring weeks or months, thereby delaying appropriate therapy and increasing the risk of treatment failure and acquired resistance.⁹ The Xpert MTB/RIF assay, a commercial nucleic acid amplification test, offers rapid TB diagnosis and simultaneous detection of rifampin resistance through real-time PCR.¹⁰ However, its detection scope is limited to a few drugs (mainly rifampin), and its performance depends on bacterial load, which can lead to false-negative or incomplete drug resistance.¹¹

Although chest X-ray (CXR) is the most commonly used and convenient first-line imaging examination for the initial screening and follow-up of pulmonary tuberculosis, its two-dimensional imaging characteristics and relatively low spatial resolution limit its ability to accurately depict the subtle structures of lung cavities, the characteristics of the cavity walls, and the heterogeneity within the lesions. In contrast, chest computed tomography (CT) with its extremely high spatial and density resolution can clearly present these key morphological details. Experienced radiologists can identify certain imaging features empirically associated with drug resistance, such as extensive multiple cavities, lung nodules, multilobar involvement, significant bronchial dissemination, and rapid disease progression.^12–15 However, visual assessment is inherently subjective, highly dependent on individual expertise, and difficult to standardize. Moreover, subtle and complex texture patterns, are challenging to discern, leading to variable and non-reproducible assessments, limiting its use as a reliable predictive tool in clinical practice. To address these limitations, there is an urgent need for an objective, quantitative, and efficient diagnostic approach that can complement microbiological testing and reduce reliance on subjective imaging interpretation. In recent years, the integration of artificial intelligence into medical imaging has led to the emergence of radiomics, an interdisciplinary technique that extracts quantitative biological information from standard medical images.¹⁶ Through high-throughput computational analysis, hundreds to thousands of features can be derived, including first-order statistical features (describing pixel value distributions), texture features (depicting pixel spatial relationships and heterogeneity), shape features, and higher-order filter features.¹⁷ These features transform visual patterns into structured data suitable for advanced modeling. Nonetheless, radiomic features may be influenced by scanning equipment, parameters, and reconstruction algorithms,¹⁸ and they primarily capture static structural information while lacking clinical context. Li et al previously developed a clinical-radiomics fusion model for predicting MDR in cavitary pulmonary TB, but it incorporated only basic clinical information such as gender and age.¹⁹ More detailed clinical data, including comorbidities, laboratory parameters, and immunological status were not considered.

Based on the current status and limitations of existing research, this study aimed to construct and validate a clinical-radiomics fusion prediction model specifically designed for cavitary pulmonary tuberculosis. By integrating more comprehensive clinical features—encompassing comorbidities, laboratory, and immunological indicators—than previous studies, along with quantitative radiomic information, and systematically comparing multiple machine learning algorithms, we sought to develop a tool that possesses high predictive performance, robust generalizability, and clinical utility. The resultant visualized nomogram emphasizes interpretability and ease of use, particularly in resource-limited settings. This model not only provides a novel approach for the early identification of drug resistance but also serves as a basis for individualized treatment decision-making, thereby advancing the practice of precision medicine in the prevention and control of tuberculosis. Furthermore, our study focuses on overall drug resistance, rather than being limited exclusively to MDR-TB. This broader scope allows for a more comprehensive assessment of resistance patterns and ensures that the findings are applicable to a wider spectrum of clinical challenges beyond MDR-TB.

Material and Methods

Patient Enrollment and Data Collection

This retrospective cohort in this study systematically collected complete clinical and CT imaging data from patients with cavitary pulmonary TB admitted to the Fourth People’s Hospital of Nanning between January 2016 and September 2025. The inclusion criteria were as follows: (1) microbiologically confirmed TB, with positive cultures from sputum, bronchoalveolar lavage fluid (BALF), or tissue biopsy and completed phenotypic DST; (2) availability of complete clinical baseline information and willingness to participate in the study; (3) no history anti-TB treatment; (4) high-quality chest CT images available; and (5) at least one well-visualized pulmonary cavity ≥1 cm on pre-treatment chest CT scans. This standard was established because the tiny cavities appearing in the images are easily confused with the cross-sections of blood vessels or small nodules, and the measurement errors associated with them are relatively large. Exclusion criteria included: (1) co-existing cavitary lung diseases (eg, cavitary lung cancer, cavitary pulmonary cryptococcosis, or pulmonary aspergillosis); (2) severe psychiatric disorders (eg, schizophrenia or major depressive disorder) or cognitive impairment, rendering patients unable to cooperate with medical history collection; (3) incomplete clinical data or CT images with severe artifacts affecting accurate delineation of cavity lesions; and (4) any prior anti-TB treatment before enrollment.

The study was approved by the Ethics Committee of the Fourth People’s Hospital of Nanning (approval number: [2025] 52) and conducted in accordance with the Helsinki Declaration. As only anonymized medical records were accessed, informed consent was waived. A total of 231 patients met the inclusion criteria. All participants were confirmed by mycobacterial culture and divided into drug-resistant TB (DR-TB, n=89) and drug-sensitive TB (DS-TB, n=142) based on DST results. DR-TB was defined as resistance to one or more anti-TB drugs, while DS-TB cases indicated full sensitivity to all first-line agents (isoniazid, rifampin, ethambutol, and pyrazinamide).

For DST, sputum and BALF samples were pretreated with N-acetyl-L-cysteine (NALC)-NaOH, vortexed, and adjusted to a final concentration of 2%. Subsequently, 100 μL of the processed suspension was incubated on agar slants and incubated for up to 8 weeks. After heat fixation, slides were stained by the Ziehl-Neelsen acid-fast method and semi-quantitatively graded (negative, +, ++, +++). Positive isolates were tested by the agar proportion method in accordance with Clinical and Laboratory Standards Institute (CLSI) guidelines. Drug susceptibility was defined as complete inhibition of bacterial growth at a given critical concentration. Patients were randomly divided into a training set (n=161) and a test set (n=70) in a 7:3 ratio. Specifically, the training set comprised 67 cases of DR-TB and 94 cases of DS-TB, while the testing set consisted of 22 cases of DR-TB and 48 cases of DS-TB. Two researchers, blinded to DST results, extracted demographic data, clinical symptoms and medical history, comorbidities, and laboratory parameters from electronic medical records.

CT Image Acquisition and Preprocessing

All CT examinations were performed using a standardized protocol using a 64-slice spiral computed tomography scanner (GE LightSpeed VCT 64, GE Healthcare, USA). Equipment calibration was conducted before each session to ensure compliance with quality control standards. Participants were positioned supine, instructed to maintain even breathing, and asked to hold their breath during scanning to minimize image artifacts.

Non-contrast chest CT was performed with the following parameters: slice thickness 5 mm, tube voltage 120 kV, and automatic tube current modulation. Images were reconstructed using both standard and high-resolution algorithms; the latter was used for radiomics analysis. All qualified images were archived in DICOM format. Before feature extraction, images were resampled to isotropic voxels (1 mm × 1 mm × 1 mm) to standardize spatial resolution. A consistent window width and level of 40 and 400, respectively, were applied to reduce noise and enhance texture feature stability.

Region of Interest (ROI) Delineation

Two radiologists (each with > 5 years of experience in chest imaging), blinded to all clinical and DST information, independently delineated the volumes of interest (VOIs) using ITK-SNAP software. Under lung window settings, cavity boundaries were manually traced slice by slice. The target lesion selection criteria were as follows: the largest or most representative cavity was chosen; if multiple cavities were present, up to the three were delineated and merged into a single VOI to capture lesion heterogeneity. Each VOI included the cavity lumen, inner wall, and adjacent consolidation or ground-glass opacity.

Radiomics Feature Extraction

Radiomic feature extraction was performed using the open-source Python package PyRadiomics. Seven categories of features were extracted: (1) first-order statistics: voxel intensity metrics, such as energy, entropy, kurtosis, and skewness; (2) shape features describing the three-dimensional morphological descriptors, such as volume, surface area, and sphericity; (3) gray-level co-occurrence matrix (GLCM) texture features describing spatial relationships between voxel pairs, quantifying image contrast, correlation, homogeneity, and energy; (4) gray-level run-length matrix (GLRLM) texture features quantifying the length of consecutive gray-level values, reflecting image roughness; (5) gray-level size zone matrix (GLSZM) texture features describing the size of connected regions in 2D images, reflecting regional uniformity; (6) neighborhood gray-tone difference matrix (NGTDM) features capturing local texture; and (7) gray-level dependence matrix (GLDM) features describing gray-level dependencies. Additionally, higher-order features were derived using Laplacian of Gaussian and wavelet filters. In total, 1,834 features were initially extracted per VOI. To assess feature reproducibility, one radiologist repeated VOI delineation for 30 randomly selected patients 4 weeks after the initial session. Intraclass correlation coefficients (ICCs) were calculated for intra- and inter-observer consistency. Features with ICC > 0.8 were retained for subsequent analysis.

Feature Selection and Radiomics Signature Construction

Features with ICC ≤ 0.8 were excluded. For the remaining features, pairwise Spearman correlation coefficients were computed, and among highly correlated pairs (|r| ≥ 0.9), the feature less correlated with the outcome variable was removed. Feature selection was then performed using least absolute shrinkage and selection operator (LASSO) regression, which applies an L1 penalty to regression coefficients to zero. The optimal penalty parameter (λ) was determined via 10-fold cross-validation based on the minimum model error. Selected features were linearly combined using their LASSO regression coefficients to compute a radiomics score (Rad-score) for each patient, representing the individual probability of drug resistance.

Predictive Model Construction and Comparison

Three predictive models were constructed and compared: (1) a clinical model using features with significant intergroup differences (P<0.05); (2) a radiomics model using only the constructed radiomics signature; and (3) a clinical-radiomics fusion model combining selected clinical predictors and the radiomics signature. Eight machine learning algorithms were compared, including logistic regression (LR), support vector machine (SVM), random forest (RF), XGBoost, and K-nearest neighbors (KNN). Hyperparameters were optimized via cross-validation and grid search to ensure model generalization and predictive performance. Model performance was quantified using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. DeLong’s test was used to compare AUCs between models. Calibration curves and the Hosmer-Lemeshow (H-L) goodness-of-fit test (P > 0.05 indicated good calibration) assessed agreement between predicted and observed outcomes. Decision curve analysis (DCA) was quantified clinical net benefit across threshold probabilities. The best-performing fusion model was visualized as a nomogram w for individualized risk prediction. Statistical analyses were performed using SPSS 23.0, R and Python.

Statistical Analysis

All statistical analyses were performed using R (version 4.4.1; R Foundation for Statistical Computing, Vienna, Austria) and Python (version 3.11.5; Python Software Foundation, Wilmington, Delaware, USA). Continuous variables were expressed as mean ± standard deviation (mean ± SD), and compared using independent-samples t-tests when normally distributed. Categorical variable was compared using Pearson’s chi-square test. Radiomic features evaluated using the Mann–Whitney U-test, and features with P < 0.05 for statistically significant.

Results

Radiomics Feature Extraction

A total of 231 patients were included and classified into the drug-resistant TB (DR-TB, n=89) and drug-sensitive TB (DS-TB, n=142) groups based on DST results. Patients were randomly divided into a training set (n=161) and a test set (n=70) in a 7:3 ratio. A total of 1,834 handcrafted features with intraclass correlation coefficient (ICC) > 0.8 were retained, ensuring good consistency (Supplementary Figure 1 A and B). All features were extracted using the PyRadiomics package (http://pyradiomics.readthedocs.io).

As shown in Supplementary Table 1, the extracted features comprised seven categories, including GLCM (440, 24.0% of the total), first-order features (360, 19.6%). GLRLM (320; 17.4%) and GLSZM (320; 17.4%), GLDM (280; 15.3%), NGTDM (100; 5.5%), and shape features (14, 0.8%). After dimensionality reduction via 10-fold cross-validation and LASSO regression, 14 optimal radiomics features were identified for subsequent modeling (Supplementary Figure 1 C and D).

Radiomics Model Construction

The predictive performance of radiomics models is summarized in Table 1. Most models achieved high discriminative ability in the training sets, with the XGBoost model achieving an AUC of 0.99. However, its performance significantly declined in the test set, suggesting potential overfitting. In contrast, the LR model maintained relatively stable and high AUC values in both sets, demonstrating better generalization. Figure 1 shows the AUC comparison for each algorithm. Based on this robustness, the LR model was selected for further fusion model development.

Table 1 Radiomics Feature Models

Figure 1 The AUC of the radiomics models. (A). Train cohort; (B). Test cohort.

Clinical Feature Selection

Baseline characteristics of patients in the training and test sets are presented in Supplementary Table 2. No significant differences were observed in age, height, weight, BMI, or immune indicators (CD3, CD4, CD8) between DS-TB and DR-TB groups (P > 0.05). However, C-reactive protein (CRP) levels were significantly higher in DR-TB patients than in DS-TB patients (training set P < 0.001, test set P = 0.003), suggesting a stronger inflammatory response. The prevalence of diabetes mellitus was also significantly higher among DR-TB patients (training set P = 0.001, test set P = 0.034), indicating a potential association between diabetes and TB drug resistance. Additionally, significant intergroup differences were observed in alanine aminotransferase (ALT) and serum calcium (Ca²⁺) levels in the training set (P < 0.05). These four clinical variables (CRP, diabetes history, ALT, and Ca²⁺) were therefore included in subsequent model construction. Supplementary Figure 2 shows the correlations between each clinical features, it indicates that Long Diameter, Short Diameter and Diameter have maximum correlation coefficient.

Clinical Model Construction

The performance of clinical models using different machine learning models is shown in Table 2. In the training set, the ExtraTrees and RandomForest models achieved high classification performance (AUCs of 1.000 (95% CI: 1.0000–1.0000) and 0.998 (95% CI: 0.9958–1.0000), respectively, with corresponding sensitivity, specificity, and F1 scores approaching 1.0. However, their performance decreased markedly in the test set, indicating overfitting. By contrast, the LR model exhibited more balanced performance, achieving AUCs of 0.735 (training) and 0.725 (test), demonstrating better generalization. Given its stability and interpretability, the LR model was selected as the base algorithm for constructing the clinical–radiomics fusion model.

Table 2 Clinical Feature Models

Construction of the Clinical-Radiomics Fusion Model

Table 3 compares the performance of different fusion models. The nomogram-based clinical-radiomics fusion model achieved the overall results, with AUCs of 0.861 in the training set and 0.884 in the test set, superior to both the Clinic Signature and Rad Signature models (Figure 2). In the test set, the Clinic Signature model had lower positive predictive value (PPV) and F1 scores (0.471 and 0.582 respectively), while the Rad Signature model achieved an AUC of 0.872, comparable to the nomogram, but with lower PPV and F1 scores. The Nomogram model exhibited high stability and strong classification performance, making it the optimal predictive model for clinical application.

Table 3 Comparison of Diagnostic Efficacy in Smear-Negative Cases

Figure 2 ROC comparison of clinical, radiomics, and fusion models. (A). Train cohort; (B). Test cohort.

Model Performance Evaluation

Model evaluation results using the H-L test and DeLong test are summarized in Supplementary Tables 3 and 4. The H-L test showed that the nomogram significantly outperformed the Clinic Signature (p = 0.004) and Rad Signature (p = 0.011) models in the training set, and remained superior to the Clinic Signature in the test set (p = 0.038). The DeLong test further validated the nomogram’s AUC was significantly higher in the test set (p = 0.035). DCA revealed that the nomogram model achieved the highest net clinical benefit across a wide range of threshold probabilities in both the training and test sets (Figure 3). Supplementary Figure 3 presents the calibration curves of the training group and the test group. This result further validates the effectiveness of the fusion model in avoiding overfitting and enhancing generalization ability by integrating multiple source features, providing a more reliable basis for clinical decision-making. The final nomogram integrates both clinical and radiomics features, allowing intuitive visualization of an individual patients’s probability of drug resistance by summing the scores of each variable along the total score axis (Figure 4). This provides clinicians with a practical tool for personalized risk assessment and treatment optimization.

Figure 3 DCA comparison of clinical, radiomics, and fusion models. (A). Train cohort; (B). Test cohort.

Figure 4 Nomogram for clinical use.

Discussion

With the deep integration of artificial intelligence into medical imaging, radiomics, an emerging interdisciplinary field, has provided a novel paradigm for precision diagnosis and treatment. In oncology, radiomics has achieved notable success in tumor classification, prognosis prediction, and treatment response assessment.^20–22 Translating this concept to TB, the abnormal immune response and the tissue destruction and repair processes triggered by drug-resistant MTB create distinct pathological changes within lung cavities and surrounding tissues. These pathological features are reflected on CT images as specific, quantifiable radiomic feature patterns. Therefore, radiomics holds great potential for identifying hidden “imaging biomarkers”, and enabling non-invasive and early prediction of drug resistance in cavitary pulmonary TB.

The radiomics model constructed in this study demonstrated certain predictive ability for distinguishing DR-TB from DS-TB. In-depth analysis of feature types provided a reliable biological and pathological perspective for interpreting the model’s decision-making mechanism. Texture features, including GLCM, GLRLM, GLSZM, GLDM, and NGTDM, collectively constituted the majority of selected features, strongly suggesting a fundamental link between TB drug resistance and intralesional heterogeneity. GLCM features, which accounted for the largest proportion (24.0%), capture the complex histopathological alterations caused by drug-resistant bacterial infections, such as the irregular interweaving of necrosis, liquefaction, and fibrosis. These microstructural disturbances manifest as complex spatial textures variations on CT imaging. Fundamental differences exist between DR-TB and DS-TB regarding host-pathogen interactions, local immune responses, and tissue repair processes, which are ultimately mapped onto the textural and structural features of CT imaging. DR-TB is frequently characterized by a delayed therapeutic response, resulting in recurrent inflammatory exudation, granulation tissue hyperplasia, and necrosis within the cavity walls, thereby leading to thicker walls and more irregular inner linings.²³ These pathologic changes manifest as specific patterns of shape and high-order filtered features on imaging; concurrently, they are reflected by increased complexity in texture metrics, such as GLCM and GLRLM parameters, specifically evidenced by elevated contrast, reduced homogeneity, and enhanced run-length non-uniformity. In contrast, shape features, which accounted for only 0.8% of all features, contributed minimally, implying that the key information for identifying drug resistance lies in the internal texture of lesions rather than their gross morphology. A comparison of eight mainstream machine learning algorithms revealed that the LR model, although not achieving the highest performance in the training set (AUC = 0.832), demonstrated superior generalization in the test set (AUC = 0.872). In contrast, complex ensemble algorithms such as XGBoost exhibited nearly perfect discrimination in the training set (AUC = 0.99) but showed substantial overfitting in the test set (AUC = 0.836). This finding suggests that for medical prediction tasks with limited sample sizes, relatively simple models (eg, LR) may outperform complex machine learning algorithms. LR offers computational efficiency, interpretability, and robust performance with small datasets, characteristics well suited to clinical applications.²⁴ Additionally, its output probabilities can be directly interpreted as odds ratios, facilitating clinical interpretation and adoption.²⁵

Regarding clinical variables, four indicators, including CRP, diabetes, ALT, and blood Ca²⁺ level, were significantly associated with TB drug resistance. These easily accessible indicators provide valuable tools for early risk assessment and offer insights into the pathophysiological mechanisms of DR-TB. Elevated CRP levels in DR-TB patients, consistent with previous findings,²⁶ reflect sustained systemic inflammation. This persistent inflammatory state may not only indicate ineffective immune system is struggling to effectively control drug-resistant bacterial infections, but may also directly contribute to tissue damage and disease progression. Moreover, this persistent inflammatory can exacerbate drug resistance by altering the pharmacokinetics of anti-TB drugs and reducing their local concentrations within lesions. Mechanistically, MTB activates the NLRP3 inflammasome in immune cells, leading the increased secretion of IL-1β and IL-18. Previous studies have shown that serum levels of these cytokines are significantly higher in patients with pulmonary TB than in healthy controls, and their concentrations are positively correlated with CRP levels.²⁷ Similar findings were also observed in our previous work.²⁸ The association between diabetes and TB drug resistance identified in this study is consistent with extensive global evidence.^29,30 Diabetes-induced immune dysregulation increases susceptibility to TB and worsens clinical outcomes. TB patients with diabetes tend to exhibit faster disease progression and higher bacterial loads.³¹ Additionally, hyperglycemia-induced intestinal barrier dysfunction and microbial dysbiosis promote systemic pathogen dissemination.³² Diabetes increases the risk of MDR-TB through multiple mechanisms, including impaired phagocytic activity, chemotactic responses, oxidative substance production, microbial proliferation, altered drug disposition, and treatment non-adherence.^33,34 These findings emphasize the need for heightened vigilance in managing TB patients with diabetes, particularly in assessing their risk of drug resistance. The observed association between elevated ALT levels and TB drug resistance may reflect hepatocellular injury resulting from prior irregular treatment or exposure to second-line anti-TB drugs.^24,35 Hepatic dysfunction can further alter drug metabolism and efficacy, potentially creating a vicious cycle that perpetuates drug resistance. In addition, abnormal serum calcium levels may indicate endocrine or metabolic disturbances associated with TB infection, although their specific mechanisms warrant further investigation.

Both single diagnostic models exhibited varying degrees of overfitting, revealing that the patterns learned from the training data did not fully generalize to new data, thereby limiting their clinical applicability. In clinical practice, diagnostic decision-making is inherently multifactorial, requiring integration of laboratory data, clinical manifestations, and imaging features. Based on this, we developed and validated a fusion model integrating clinical features and radiomic features for predicting drug resistance in cavitary pulmonary TB. Compared with the individual clinical or radiomic models, the fusion model exhibited superior predictive performance and clinical utility in distinguishing DR-TB from DS-TB. This confirms the advantages of multimodal data integration and provides a novel non-invasive approach for early identification of drug resistance. The clinical-radiomics fusion model (nomogram) maintained stable and excellent predictive performance in both the training and test sets (AUC = 0.861 and 0.884, respectively). Notably, its slightly higher performance in the test set suggests good generalization without significant overfitting. In contrast, the single clinical model and radiomics achieved an AUC of 0.725 and 0.872, respectively. These results clearly demonstrate a synergistic effect, where combining clinical and imaging features yields greater predictive power than either modality alone. From a translational perspective, decision curve analysis showed that the fusion model provided the highest net clinical benefit across a range of threshold probabilities. Furthermore, visualization of the model as a nomogram enhances its clinical applicability, enabling rapid, individualized estimation of drug resistance risk to support personalized treatment decisions. The clinical-radiomics fusion model constructed in this study differs in design philosophy from a recent study that utilized large-scale samples and deep learning architectures for the precise diagnosis and typing of drug-resistant tuberculosis.³⁶ Supported by sufficient data, that deep learning study achieved excellent predictive accuracy, representing a “high-precision” paradigm tailored for large-scale medical centers. In contrast, the core advantages of the present study lie in the parsimony, interpretability, and strong adaptability to small-to-medium-sized samples of its model. This linear model, based on clinical and radiomic features, significantly reduces reliance on large-scale annotated data and high-performance computing power; furthermore, its output (eg, nomograms) is more easily understood and trusted by clinicians. Therefore, rather than aiming to outperform complex deep learning models in terms of performance, this model seeks to provide a practical and easily deployable decision support tool for primary healthcare institutions, which are often characterized by limited computational resources and small sample sizes, thereby filling a gap in current research regarding “accessibility” and “practicality.”

Although the fusion model constructed in this study demonstrates promising predictive performance and clinical utility, several limitations remain. First, this study was designed as a single-center retrospective analysis; therefore, its generalizability requires further validation through multi-center, prospective external data. Second, this study is essentially a radiological association analysis. The established model lacks direct correlation with histopathology, precluding a definitive elucidation of the causal mechanisms linking radiomic features to specific pathological changes in DR-TB (eg, granuloma structure, necrosis patterns). Although the logistic regression-based parsimonious model enhances applicability in primary healthcare settings, its capacity to handle complex non-linear relationships may be inferior to that of deep learning models, which inherently restricts its performance ceiling in ultra-large-scale and highly heterogeneous datasets. Furthermore, the inability to achieve precise subtyping of drug resistance profiles limits the direct clinical value of this study in guiding individualized precision treatment. Future research should focus on conducting multi-center external validation and integrating histopathological specimens or molecular imaging techniques to deeply elucidate the biological basis of imaging features. Concurrently, it is necessary to explore strategies to optimize model complexity and generalizability using large-scale data and to develop multi-class models for precise resistance subtyping, thereby enhancing the clinical value of radiomic features in guiding individualized precision therapy. It must be emphasized that the clinical-radiomics fusion model developed in this study is intended to serve as an auxiliary support tool for clinical decision-making to facilitate early risk stratification, and should not be considered a complete replacement for conventional standard drug susceptibility testing.

Conclusion

In summary, this study successfully developed and validated a clinical-radiomics fusion model that demonstrated excellent performance and clinical utility in predicting drug resistance in cavitary pulmonary TB. The results highlight the advantages of multimodal data fusion in medical prediction modeling and establish a novel, non-invasive methodological framework for early drug resistance assessment. By transforming complex computational algorithms into an intuitive, user-friendly nomogram, this study effectively bridged the gap between artificial intelligence and clinical decision-making. The proposed model provides a powerful tool to support individualized treatment planning and precision medicine in TB management. Future research should focus on external validation, across diverse populations, continuous algorithm optimization, and integration into clinical workflows to ensure robust, real-world applicability and maximize patient benefit.