Abstract
Background
Cerebral tuberculomas and brain metastases of bronchial carcinoma are often indistinguishable on magnetic resonance imaging (MRI), leading to misdiagnosis. This study develops a contrast-enhanced MRI radiomics-based nomogram to accurately differentiate them.
Methods
A total of 482 patients (290 males, 192 females; mean age 56.08±12.96 years), including 254 with cerebral tuberculomas and 228 with brain metastases of bronchial carcinoma were enrolled from two centers. Radiomic features were extracted, standardized, and reduced via significance testing, Pearson correlation, and least absolute shrinkage and selection operator (LASSO) regression. A radiomic model was built, then integrated with clinical predictors into a nomogram. Performance was assessed across training, internal validation, and external validation cohorts using receiver operating characteristic (ROC) curves, calibration, and decision curve analysis (DCA).
Results
A total of 13 radiomic features were selected for model development. The radiomic model demonstrated areas under the curve (AUCs) of 0.907, 0.891 and 0.886 in training, internal validation, and independent external validation cohorts, respectively, significantly outperforming models based solely on clinical factors, which achieved AUCs of 0.781, 0.712, and 0.828 (P<0.05). The nomogram, which integrated the radiomic scores (Rad-scores), age, ring enhancement, meningeal thickening, and carcinoembryonic antigen (CEA) levels, exhibited superior performance with AUCs of 0.951, 0.937, and 0.940 across the three cohorts. DCA demonstrated that the nomogram provided greater clinical net benefits across a range of threshold probabilities, and calibration curves indicated strong consistency between predicted probabilities and actual outcomes.
Conclusions
A nomogram combining clinical factors and contrast-enhanced MRI radiomics accurately differentiates cerebral tuberculomas from brain metastases of bronchial carcinoma, offering robust decision support for individualized diagnosis and treatment.
Keywords: Cerebral tuberculomas, brain metastases of bronchial carcinoma, brain, magnetic resonance imaging (MRI), radiomics
Introduction
Cerebral tuberculomas and brain metastases of bronchial carcinoma are pathologically distinct—one arising from Mycobacterium tuberculosis infection, the other from hematogenous spread of solid tumors (1,2). Nonetheless, both diseases can cause severe neurological impairment. Cerebral tuberculomas is the main clinical manifestation of central nervous system (CNS) tuberculosis, accounting for 10–30% of all such cases (3-5). Meanwhile, lung cancer—a highly prevalent and lethal malignancy globally—leads to brain metastases in 20–65% of patients as survival extends (6). Clinically and radiologically, they present with similar symptoms (headaches, seizures) and nonspecific imaging features, making reliable non-invasive differentiation difficult. However, their treatment and prognosis diverge sharply, underscoring the critical need for accurate preoperative distinction (7,8). Moreover, differentiating brain metastases of bronchial carcinoma from cerebral tuberculomas is clinically more valuable than distinguishing tuberculoma from other infectious or inflammatory lesions. First, the two entities differ substantially in treatment, prognosis, and diagnostic pathways; misdiagnosis can result in critical clinical errors. Second, since most other inflammatory or rare infectious lesions are benign, misdiagnosis involving them poses relatively limited clinical risk and management deviation. Additionally, although multifocal glioblastoma and primary CNS lymphoma are important differential diagnoses of multifocal intracranial lesions, these conditions were excluded from the present study due to their relatively low incidence and the specific scope of our investigation, focusing on more common etiologies (9,10).
Conventional magnetic resonance imaging (MRI) remains essential for evaluating intracranial lesions, relying on features such as signal intensity, morphology, margins, and contrast enhancement patterns. However, both tuberculomas and metastases frequently present as round or irregular nodules with overlapping nodular or ring-like enhancement (11,12), limiting diagnostic specificity. Such nonspecific imaging features lead to subjective radiological interpretation, frequently requiring invasive examinations to achieve a definitive pathological diagnosis. However, stereotactic biopsy for deep intracranial lesions carries notable surgical risks and technical challenges, limiting its routine clinical application. Second, lumbar puncture (LP) demonstrates limited diagnostic yield, especially in patients with isolated cerebral tuberculomas, as previously reported (13). These limitations underscore the urgent need for non-invasive, objective diagnostic tools.
Radiomics enables non-invasive detection of subtle pathological changes invisible to the naked eye and has proven valuable in tumor grading and infectious CNS diseases (14,15). However, individual radiomic features are often unstable due to technical variability, while clinical parameters alone offer limited predictive power (16-18). Integrating both data types has therefore emerged as a promising strategy, yet comprehensive models specifically designed to differentiate cerebral tuberculomas from brain metastases of bronchial carcinoma remain lacking (19). In this study, two representative clinical biomarkers, carcinoembryonic antigen (CEA) and interferon-γ release assays (IGRAs), were integrated for auxiliary discrimination. Previous clinical data have demonstrated that CEA exhibits a diagnostic sensitivity of 65–80% and specificity of 85–95% for intracranial lesions secondary to malignancies, whereas IGRAs exhibit a sensitivity of 70–90% and specificity of 90–98% for tuberculosis diagnosis (20,21). Accordingly, we incorporated these two biomarkers into our predictive model to investigate their potential to further enhance the overall differential diagnostic accuracy for the target intracranial lesions.
We aim to develop and validate a nomogram that integrates contrast-enhanced MRI radiomic features with key clinical indicators using multi-center data. Through rigorous feature selection, model optimization, and validation across internal and independent external cohorts, we seek to establish a robust, generalizable tool for early and accurate differential diagnosis—supporting individualized treatment decisions without recourse to invasive procedures. We present this article in accordance with the STARD reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1518/rc).
Methods
Study population
This retrospective study was approved by the Ethics Committees of Guangzhou Chest Hospital {No. Xiongyi Ethics [2023] No. 17} and Hunan Chest Hospital (No. LS2026011201). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments, and individual consent for this retrospective analysis was waived. The initial cohort included patients who underwent radiological assessments at both institutions from January 2021 to December 2025. Inclusion criteria were: (I) meeting the diagnostic criteria for cerebral tuberculomas or brain metastases of bronchial carcinoma, with complete clinical and contrast-enhanced MRI data; (II) possessing at least one identifiable lesion in the brain parenchyma on contrast-enhanced MRI scans; (III) a definite diagnosis confirmed either by histopathological examination of samples obtained via resection or biopsy, or by confirmed molecular diagnostic findings. Patients were excluded based on the following criteria: (I) unclear clinical diagnosis, with ineffective anti-tuberculosis or anti-tumor treatments; (II) unclear pathological diagnosis attributed to insufficient tissue samples; (III) history of other intracranial malignancies or co-infections with other pathogens (e.g., viral, bacterial, or fungal infections); and (IV) poor image quality due to motion artifacts. The reference standard for the final diagnosis of cerebral tuberculomas or brain metastases of bronchial carcinoma was defined by one of the following criteria: histopathological examination of surgically resected or biopsied tissue, molecular analysis of cerebrospinal fluid (CSF) samples, or confirmed shrinkage and absorption of intracerebral lesions after standardized anti-tuberculosis or anti-tumor therapy (≥50% reduction in lesion size on follow-up MRI). Contrast-enhanced MRI and the reference standard procedures (biopsy or CSF collection) were performed during the same hospital admission, with no intervening antituberculosis or antineoplastic therapy administered prior to pathological confirmation. The specific interval varied according to clinical scheduling but remained within a 2-week window. For patients for whom the reference standard was based on treatment response (≥50% reduction in lesion size on follow-up MRI), the time interval between the baseline contrast-enhanced MRI and the follow-up MRI was fixed at 3 months post-initiation of standardized anti-tuberculosis or anti-tumor therapy.
Ultimately, the study included 482 patients, comprising 406 from Guangzhou Chest Hospital (204 diagnosed with cerebral tuberculomas and 202 with brain metastases of bronchial carcinoma), randomly stratified into a training set (n=284) and an internal validation set (n=122), all brain metastases in this cohort arose from primary lung cancer, with histological subtypes comprising squamous cell carcinoma (n=77, 38.12%), adenocarcinoma (n=86, 42.57%), large cell carcinoma (n=33, 16.34%), and small cell carcinoma (n=6, 2.97%). A total of 76 patients from Hunan Chest Hospital were allocated to the independent external validation cohort (n=76). A detailed workflow diagram is presented in Figure 1.
Figure 1.
Participant flow diagram of the study. This flowchart illustrates the selection process of patients meeting all prespecified inclusion criteria. Cases excluded prior to eligibility assessment lacked consistent documentation owing to the retrospective nature of data collection. MRI, magnetic resonance imaging.
Image acquisition and analysis
MRI scans were acquired using two 1.5 T platforms: 406 patients were imaged on the UMR560 system (United Imaging) at Guangzhou Chest Hospital, and 76 patients were imaged on the Siemens Magnetom Avanto 1.5 T scanner at Hunan Chest Hospital. Identical fat-saturated contrast-enhanced three-dimensional (3D) T1-weighted sequences were applied to all patients. Before scanning, patients were screened to ensure the absence of metallic objects within their bodies. Additionally, individuals with known allergies to contrast agents were excluded from the contrast-enhanced scan. The contrast agent, gadopentetate dimeglumine, was administered via superficial or deep veins in the forearm prior to imaging. The contrast dose was determined based on the patient’s weight, typically ranging from 0.1 to 0.2 mmol/kg, with an injection rate of 3–4 mL/s. Following this, axial T1-weighted gradient echo 3D fat suppression scans were performed, with parameters set as follows: slice thickness of 2 mm, slice gap of 0 mm, TR of 11.5 ms, TE of 4.8 ms, image reconstruction matrix size of 348×512 pixels, number of excitations of 2, flip angle of 20°, and excitation frequency of 1. Two senior radiologists with more than 10 years of experience in cranial disease diagnosis independently reviewed the MRI images in a double-blinded manner.
Image segmentation and feature extraction
Intra- and inter-observer reproducibility was assessed using intraclass correlation coefficients (ICCs). A representative random sample of 60 contrast-enhanced MRI images (30 cerebral tuberculomas and 30 brain metastases), accounting for approximately 12.4% of the total dataset, was selected for ICC testing. Two radiologists, masked to all clinical data, laboratory values, and final histopathological diagnoses, independently delineated regions of interest (ROIs) on a slice-by-slice basis to construct whole-lesion volumes of interest (VOIs). Radiologist 1 repeated the segmentation after a 2-week interval. An ICC threshold >0.80 was predefined as indicating good consistency. Following the reliability assessment, Radiologist 1 performed VOIs segmentation for the remaining 422 cases, leveraging the excellent reproducibility confirmed in the representative sample. Only radiomic features with ICC >0.80 were retained for subsequent analysis. For patients with multiple intracranial lesions, a single representative lesion was selected for manual segmentation and radiomic feature extraction. All statistical analyses were performed on a per-patient basis to mitigate clustering bias arising from multiple lesions within the same patient.
Using 3D Slicer software (version 5.4.0, https://www.slicer.org/), using contrast-enhanced MRI, ROIs were delineated on each slice along the interface between each enhancing lesion and adjacent brain parenchyma to generate whole-lesion VOIs encompassing the entire enhancing component. Reproducibility validation was also performed based on these whole-lesion VOIs, ensuring methodological consistency. Representative segmentation examples are shown in Figure 2. From each VOIs, 1,746 candidate radiomic features were extracted using the ImferScholar platform (Version 6.7), including first-order statistics, gray level size zone matrix (GLSZM), neighboring gray tone difference matrix (NGTDM), and gray level dependence matrix (GLDM) features (Table S1). Dimensionality reduction and feature selection were performed strictly within the training set using a nested 10-fold cross-validation (CV) loop to prevent data leakage. Specifically, in each outer CV fold, we first applied univariate significance testing (P<0.05) on the training sub-fold to exclude features unrelated to the outcome. Next, Spearman correlation analysis (|ρ|>0.85) was used to remove highly redundant features. Then, within the same training sub-fold, least absolute shrinkage and selection operator (LASSO) regression with inner 10-fold CV was applied to select the most informative features. The selected features from each outer fold were used to train the model on that fold. The final optimal subset of 13 radiomic features was determined as those that appeared most frequently across all outer CV folds. These 13 features were then linearly combined using their respective LASSO coefficients to generate a radiomic score (Rad-score) for each patient, thereby establishing the radiomic model. The index test was contrast-enhanced MRI radiomics analysis, which involved extracting radiomic features from preprocessed images.
Figure 2.
Representative MRI findings of cerebral tuberculomas and brain metastases of bronchial carcinoma. (A,B) MRI images obtained by scanning the perituberculoma region and manually delineated ROI. (C,D) MRI images obtained by scanning the perimetastatic region and manually delineated ROI. “#1” and “#21” represent the label numbers of ROIs manually delineated by the physician. MRI, magnetic resonance imaging; ROI, region of interest.
Model construction, evaluation, and nomogram development
We developed clinical, radiomic, and combined models for distinguishing cerebral tuberculomas from brain metastases of bronchial carcinoma. Following feature selection, the radiomic model was constructed using the ImferScholar platform (version 6.7). The workflow for feature extraction and model development is illustrated in Figure 3. For the clinical model, univariate logistic regression was used to identify significant risk factors and then select independent predictors via multivariate analysis. Meningeal thickening in this study denotes leptomeningeal (pia-arachnoid) thickening. No patient received LP within 7 days before MRI, and leptomeningeal changes are not caused by LP. Significant variables from both univariate and multivariate analyses were subsequently incorporated into the combined model. A nomogram was developed to visualize the combined model, enabling assessment of variable contributions and facilitating prediction of individual risk.
Figure 3.
Selection of radiomics features using the LASSO algorithm. (A) Radiomic feature coefficient trajectories during LASSO feature selection. (B) Regularization path of binomial deviance in LASSO regression. In the model, optimization of omics parameters ultimately retained the final set containing 13 radiomics features. (C) Bar chart of radiomics feature coefficients: the Y-axis represents the selected 13 radiomics features, and the X-axis represents the coefficients of radiomics features. 3D, three-dimensional; LASSO, least absolute shrinkage and selection operator.
Model discrimination was compared using the DeLong test [area under the curve (AUC), sensitivity, specificity, accuracy] across all cohorts. Nomogram performance was further assessed via calibration (Hosmer-Lemeshow) and decision curve analyses.
Statistical analysis
Statistical analyses were performed using R software (version 4.4.2). Throughout the analysis, various specialized statistical packages—including glmnet, cmprsk, rms, rmda, and devtools—were employed. Continuous data were presented as mean ± standard deviation. Univariate and multivariate logistic regression models were utilized to identify independent predictors of outcomes based on collected clinical variables, with P<0.05 considered statistically significant. The discriminative performance of each model was appraised through the construction of receiver operating characteristic (ROC) curves. Statistical inference was conducted via two-tailed tests, with significance set at P<0.05. The sample size was determined using a power of 80% and an alpha of 0.05 to detect a minimum difference in the AUC of 0.1 between models. Author Liang served as the statistical guarantor, responsible for the transparency of statistical analyses.
Missing data were addressed via complete-case analysis (listwise deletion), as the proportion of missing values for key clinical variables (CEA, IGRAs, and imaging features) was less than 5% across the entire cohort. For the radiomic features, there were no missing values, as features were consistently extracted from all available VOIs. Sensitivity analyses comparing results with and without imputation were not performed, as the low proportion of missing data was unlikely to affect the overall findings.
By following these methodological approaches, we aim to develop a robust and interpretable model capable of distinguishing between cerebral tuberculomas and brain metastases of bronchial carcinoma, thus enhancing the diagnostic accuracy and facilitating personalized treatment strategies for patients.
Results
Baseline characteristics of patients
This study analyzed 482 patients (290 males, 192 females; mean age 56.08±12.96 years), including 254 with cerebral tuberculomas and 228 with brain metastases of bronchial carcinoma. Of these, 406 patients were from Guangzhou Chest Hospital (204 tuberculomas, 202 metastases), and 76 from Hunan Chest Hospital (50 tuberculomas, 26 metastases). The Guangzhou cohort was randomly split into a training set (142 tuberculomas, 142 metastases) and an internal validation set (62 tuberculomas, 60 metastases). The Hunan cohort served as an independent external validation set (50 tuberculomas, 26 metastases). In a cohort of 228 patients with brain metastases, histological subtypes included adenocarcinoma (n=103, 45.18%), squamous cell carcinoma (n=87, 38.16%), large cell carcinoma (n=32, 14.04%), and small cell carcinoma (n=6, 2.63%). Solitary lesions were observed in 63 patients (27.63%) and multiple lesions in 165 patients (72.37%). Table 1 summarizes the demographic and clinical imaging characteristics of the study population.
Table 1. Statistical characteristics of clinical and imaging parameters in patients with cerebral tuberculomas and brain metastases of bronchial carcinoma.
| Characteristic | Training cohort (n=284) | Internal validation cohort (n=122) | External validation cohort (n=76) | |||
|---|---|---|---|---|---|---|
| Cerebral tuberculomas (n=142) | Brain metastases of bronchial carcinoma (n=142) | Cerebral tuberculomas (n=62) | Brain metastases of bronchial carcinoma (n=60) | Cerebral tuberculomas (n=50) | Brain metastases of bronchial carcinoma (n=26) | |
| Sex | ||||||
| Male | 84 (59.15) | 86 (60.56) | 33 (53.23) | 40 (66.67) | 28 (56.00) | 19 (73.08) |
| Female | 58 (40.85) | 56 (39.44) | 29 (46.77) | 20 (33.33) | 22 (44.00) | 7 (26.92) |
| Age (years) | 53.94±16.65 | 57.85±11.54 | 57.73±13.12 | 60.62±8.98 | 48.08±17.52 | 59.00±11.25 |
| Emphysema | ||||||
| Present | 35 (24.65) | 43 (30.28) | 15 (24.19) | 24 (40.00) | 12 (24.00) | 8 (30.77) |
| Absence | 107 (75.35) | 99 (69.72) | 47 (75.81) | 36 (60.00) | 38 (76.00) | 18 (69.23) |
| Diabetes mellitus | ||||||
| Present | 38 (26.76) | 31 (21.83) | 20 (32.26) | 16 (26.67) | 13 (26.00) | 5 (19.23) |
| Absence | 104 (73.24) | 111 (78.17) | 42 (67.74) | 44 (73.33) | 37 (74.00) | 21 (80.77) |
| Whole-nodule enhancement | ||||||
| Present | 98 (69.01) | 92 (64.79) | 46 (74.19) | 37 (61.67) | 30 (60.00) | 20 (76.92) |
| Absence | 44 (30.99) | 50 (35.21) | 16 (25.81) | 23 (38.33) | 20 (40.00) | 6 (23.08) |
| Ring enhancement | ||||||
| Present | 105 (73.94) | 82 (57.75) | 45 (72.58) | 38 (63.33) | 39 (78.00) | 15 (57.69) |
| Absence | 37 (26.06) | 60 (42.25) | 17 (27.42) | 22 (36.67) | 11 (22.00) | 11 (42.31) |
| Meningeal thickening | ||||||
| Present | 71 (50.00) | 52 (36.62) | 32 (51.61) | 19 (31.67) | 29 (58.00) | 8 (30.77) |
| Absence | 71 (50.00) | 90 (63.38) | 30 (48.39) | 41 (68.33) | 21 (42.00) | 18 (69.23) |
| Hydrocephalus | ||||||
| Present | 48 (33.80) | 37 (26.06) | 19 (30.65) | 14 (23.33) | 17 (34.00) | 8 (30.77) |
| Absence | 94 (66.20) | 105 (73.94) | 43 (69.35) | 46 (76.67) | 33 (66.00) | 18 (69.23) |
| CA125 (U/mL) | 45.75 (34.55,123.58) | 49.57 (25.66,101.73) | 64.33 (34.55,112.37) | 38.24 (24.63,87.02) | 40.07 (33.67,101.38) | 33.04 (24.58,56.59) |
| CEA (ng/mL) | 4.30 (3.23,5.73) | 25.67 (4.59,70.16) | 4.23 (3.21,5.06) | 8.32 (4.26,29.20) | 4.13 (3.21,5.06) | 16.36 (4.68,68.31) |
| CA199 (U/mL) | 33.76 (5.47,56.66) | 43.56 (23.56,54.73) | 35.55 (5.72,65.29) | 36.23 (11.29,58.58) | 29.45 (4.70,56.66) | 39.62 (18.56,54.19) |
| IGRAs | ||||||
| Present | 55 (38.73) | 37 (26.06) | 22 (35.48) | 16 (26.67) | 19 (38.00) | 8 (30.77) |
| Absence | 87 (61.27) | 105 (73.94) | 40 (64.52) | 44 (73.33) | 31 (62.00) | 18 (69.23) |
Continuous variables are expressed as median (interquartile range). Other data are presented as patient numbers, with percentages in parentheses or mean ± standard deviation. CA125, carbohydrate antigen 125; CA199, carbohydrate antigen 199; CEA, carcinoembryonic antigen; IGRA, interferon-γ release assay.
Reproducibility of feature extraction
The process of radiomic feature extraction demonstrated robust reproducibility. The intra-observer ICCs ranged from 0.823 to 0.935, while the inter-observer ICCs fell between 0.802 and 0.896, with all values surpassing the threshold of >0.80, signifying stable and dependable feature extraction results.
Feature selection and rad-score distribution
Our analysis identified a total of 1,746 radiomic features extracted from each region of interest (VOIs) on contrast-enhanced MRI images. These features encompassed first-order statistics, GLSZM, NGTDM, GLDM (refer to Table S1). All extracted radiomic features underwent preprocessing and standardization via the Z-score method. Through rigorous significance testing, we selected features that achieved statistical significance (P<0.05). Furthermore, we eliminated features exhibiting a Pearson correlation coefficient greater than 0.85 between the conditions, narrowing down to 237 radiomic features. We then employed LASSO regression to achieve dimensionality reduction, ultimately isolating a subset of 13 predictive radiomic features to construct our radiomic model (as illustrated in Figure 3A-3C). The radiomic feature label was constructed by linearly combining these features, weighted according to their respective coefficients (Figure 4 presents the waterfall plot of Rad-score). The waterfall plot effectively visualizes the Rad-score distribution characteristics for cerebral tuberculomas and brain metastases of bronchial carcinoma across the training, internal validation, and external validation cohorts.
Figure 4.
Waterfall plots of Rad_score across different datasets. (A) Waterfall plot of Rad_score for the training set. (B) Waterfall plot of Rad_score for the internal validation set. (C) Waterfall plot of Rad_score for the independent external validation set. The waterfall plot is ordered in descending order of Rad_score. Rad_score, radiomic score.
Construction of clinical model, radiomic model, and combined model
Statistical analysis, specifically the Wilcoxon test, revealed significant differences in Rad-score between the cerebral tuberculomas and brain metastases of bronchial carcinoma groups (P<0.05). In contrast, significant differences were detected in factors such as age, ring enhancement, meningeal thickening, and levels of biological markers [CEA, cancer antigen 19-9 (CA199), IGRAs] (P<0.05). Through multivariate regression analysis, we identified age (P=0.002), CEA (P<0.001), meningeal thickening (P=0.002), ring enhancement (P=0.02), and IGRAs (P=0.02) as independent predictive factors for lesion classification, leading us to incorporate these variables into the clinical model (Table 2). Ultimately, we constructed a combined model integrating both Rad-score and the independent predictive factors identified.
Table 2. Univariate and multivariate logistic regression analysis.
| Variable | Univariable analysis | Multivariable analysis (clinical model parameters) | Multivariable analysis (combined model parameters) | |||
|---|---|---|---|---|---|---|
| OR (95% CI) | P value | OR (95% CI) | P value | OR (95% CI) | P value | |
| Sex | 0.943 (0.586–1.516) | 0.81 | – | – | – | – |
| Age (years) | 1.027 (1.007–1.048) | 0.01 | 1.038 (1.014–1.062) | 0.002 | 1.041 (1.007–1.076) | 0.02 |
| Emphysema | 1.328 (0.788–2.249) | 0.29 | – | – | – | – |
| Diabetes mellitus | 0.764 (0.441–1.316) | 0.33 | – | – | – | – |
| Whole-nodule enhancement | 0.826 (0.502–1.355) | 0.45 | – | – | – | – |
| Ring enhancement | 0.482 (0.29–0.792) | 0.004 | 0.504 (0.286–0.891) | 0.02 | 0.356 (0.148–0.858) | 0.02 |
| Meningeal thickening | 0.578 (0.358–0.926) | 0.02 | 0.416 (0.237–0.730) | 0.002 | 0.292 (0.125–0.686) | 0.01 |
| Hydrocephalus | 0.69 (0.412–1.148) | 0.16 | – | – | – | – |
| CA125 (U/mL) | 0.999 (0.996–1.001) | 0.30 | – | – | – | – |
| CEA (ng/mL) | 1.017 (1.011–1.026) | <0.001 | 1.021 (1.013–1.030) | <0.001 | 1.030 (1.017–1.042) | <0.001 |
| CA199 (U/mL) | 1.005 (1.002–1.009) | 0.003 | – | – | – | – |
| Number of lesions (n) | ||||||
| >10 | 1.109 (0.662–1.861) | 0.70 | – | – | – | – |
| 1 | 1.776 (0.844–3.887) | 0.14 | – | – | – | – |
| 2~10 | 1.602 (0.995–2.593) | 0.05 | – | – | – | – |
| IGRAs | 0.557 (0.335–0.92) | 0.02 | 0.508 (0.283–0.911) | 0.02 | – | – |
| Rad-score | 0.893 (0.867–0.915) | <0.001 | – | – | 0.876 (0.846–0.906) | <0.001 |
Reasons for excluding IGRAs in the joint model: collinearity with Rad-score (VIF >10), resulting in a reduced AIC value after exclusion. AIC, Akaike information criterion; CA125, cancer antigen 125; CA199, carbohydrate antigen 19-9; CEA, carcinoembryonic antigen; CI, confidence interval; IGRA, interferon-γ release assay; OR, odds ratio; Rad-score, radiomic score; VIF, variance inflation factor.
Performance of clinical model, radiomic model, and combined model
The ROC curves were used to evaluate the performance of the clinical, radiomic, and combined models across the training, internal validation, and external validation cohorts. The AUC, sensitivity, specificity, and accuracy for each model are summarized in Figure 5 and Table 3.
Figure 5.
ROC curves of radiomics model, clinical model, and combined model in predicting cerebral tuberculomas and brain metastases of bronchial carcinoma in the training cohort (A), internal validation cohort (B), and external validation cohort (C). AUC, area under the curve; CI, confidence interval; ROC, receiver operating characteristic.
Table 3. Comparison of diagnostic performance of clinical models, radiomics models, and combined models in training cohorts, internal validation cohorts, and external validation cohorts.
| Model | AUC (95% CI) | Accuracy (%) | Sensitivity (%) | Specificity (%) | PPV (%) | NPV (%) |
|---|---|---|---|---|---|---|
| Clinical model | ||||||
| Training cohort | 0.781 (0.729–0.834) | 71.83 | 67.61 | 76.06 | 73.85 | 70.13 |
| Internal validation cohort | 0.712 (0.621–0.802) | 66.39 | 55.00 | 77.42 | 70.21 | 64.00 |
| External validation cohort | 0.828 (0.735–0.920) | 76.32 | 76.92 | 76.00 | 62.50 | 86.36 |
| Radiomics model | ||||||
| Training cohort | 0.907 (0.873–0.941) | 84.86 | 79.58 | 90.14 | 88.98 | 81.53 |
| Internal validation cohort | 0.891 (0.836–0.947) | 81.97 | 93.33 | 70.97 | 75.68 | 91.67 |
| External validation cohort | 0.886 (0.813–0.959) | 76.32 | 92.31 | 68.00 | 60.00 | 94.44 |
| Combined model | ||||||
| Training cohort | 0.951 (0.929–0.973) | 88.03 | 84.51 | 91.55 | 90.91 | 85.53 |
| Internal validation cohort | 0.937 (0.894–0.98) | 90.16 | 86.67 | 93.55 | 92.86 | 87.88 |
| External validation cohort | 0.940 (0.890–0.989) | 86.84 | 80.77 | 90.00 | 80.77 | 90.00 |
AUC, area under the curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value.
All three models demonstrated good discriminative ability. The clinical model achieved AUCs of 0.781, 0.712, and 0.828 in the training, internal validation, and external validation cohorts, respectively. The radiomic model showed excellent performance, with corresponding AUCs of 0.907, 0.891, and 0.886. The combined model outperformed both, achieving AUCs of 0.951, 0.937, and 0.940 across the three cohorts.
The DeLong test confirmed that the combined model’s AUC was significantly higher than those of the clinical and radiomic models in all cohorts (P<0.05). Significant differences were also observed between the clinical and radiomic models in the training cohort (0.781 vs. 0.907, P<0.001) and internal validation cohort (0.712 vs. 0.891, P=0.002), as well as between the clinical and combined models in the external validation cohort (0.828 vs. 0.937, P=0.02).
Nomogram and its validation
A nomogram was developed to visualize the combined model and its contributing variables (Figure 6A). Calibration, assessed by the Hosmer-Lemeshow test, showed no significant deviation from ideal predictions in the training P=0.79 internal validation P=0.20, or external validation P=0.28 cohorts (Figure 6B). Decision curve analysis (DCA) indicated that the nomogram offers a high clinical net benefit for differential diagnosis across threshold probabilities from 0.1 to 0.9 (Figure 6C).
Figure 6.
Nomogram and calibration curve model evaluation for distinguishing cerebral tuberculomas from brain metastases of bronchial carcinoma based on training cohort. (A) Nomogram of a combined model based on clinical indicators and radiomics scores. “1” represents presence; “0” represents absence. (B) Calibration curve: Evaluating the consistency between the predicted probabilities from the nomogram and actual clinical observations in training cohorts, internal validation cohorts, and external validation cohorts. (C) DCA of each model for classifying cerebral tuberculomas versus brain metastases of bronchial carcinoma. The x-axis represents threshold probabilities, and the y-axis represents net benefit. The blue curve represents the combined model, the green curve represents the radiomics model, and the yellow curve represents the clinical model. CEA, carcinoembryonic antigen; DCA, decision curve analysis; IGRA, interferon-γ release assay; Rad_score, radiomic score.
Discussion
In this multi-center study, we developed and externally validated a nomogram integrating contrast-enhanced MRI radiomic signatures and clinical variables to distinguish cerebral tuberculomas from brain metastases of bronchial carcinoma. The established model exhibited robust discrimination and calibration performance across cohorts. Given the markedly distinct treatment regimens and prognoses for these two intracranial lesions, this non-invasive quantitative tool offers significant clinical utility for facilitating individualized differential diagnosis and guiding clinical decision-making.
We identified 13 core radiomic features for further analysis, encompassing a variety of categories such as wavelet transforms, first-order statistics, gray-level co-occurrence matrices (GLCM), and gray-level run length matrices (GLRLM), among others. These features elucidate the quantitative morphological differences between tuberculoma and brain metastases of bronchial carcinoma on MRI, particularly highlighting distinctions in tumor boundary smoothness and tissue density distributions. Existing literature suggests that tuberculomas, as a result of granulomatous inflammation induced by Mycobacterium tuberculosis, typically display relatively well-defined margins and ring enhancement. This characteristic effectively distinguishes them from the infiltrative boundaries that are often observed in various malignant tumors (1,22). Cerebral tuberculomas are characterized by smooth and homogeneous ring-enhancing rims. Whereas brain metastases of bronchial carcinoma exhibit heterogeneous ring enhancement with intramural nodules or irregular inner margins, these lesions generally display marked cellular heterogeneity and angiogenesis, manifesting as blurred margins and heterogeneous internal enhancement patterns (22). Additionally, Cerebral tuberculomas and brain metastases of bronchial carcinoma typically exhibit closed, complete ring enhancement, whereas open, incomplete ring enhancement is a typical sign of inflammatory demyelinating lesions (23).
Mechanistically, our findings indicate that the radiomic features we analyzed reflect notable differences in the tissue microenvironment associated with each condition. The pathological characteristics inherent to tuberculomas lead to first-order features exhibiting lower overall density and a more symmetrical gray-level distribution. Additionally, GLCM features reveal an enhanced texture uniformity, while shape features indicate greater flatness due to the tumors’ regular morphologies. On the other hand, brain metastases of bronchial carcinoma comprise densely populated cancerous regions and areas of necrosis, characterized by a rich vascular supply and uneven necrotic patches. Contrast-enhanced MRI often reveals a heterogeneous ring enhancement consistent with significant pathological variability. The corresponding radiomic features associated with brain metastases of bronchial carcinoma include first-order statistics showing higher mean values due to abundant tumor blood supply, coupled with positive skewness in regions dense with cancer cells. Furthermore, GLCM features exhibit lower maximum correlation coefficients, while wavelet transform features identify high-frequency variability at the tumor infiltration edges (24,25).
This study further delineates the capacity of high-dimensional feature analysis to uncover distinctions that extend beyond traditional imaging observations. Advanced parameters such as GLCM and texture heterogeneity effectively differentiate between cerebral tuberculomas and brain metastases of bronchial carcinoma. Prior analyses generally depended on subjective image evaluation and basic low-dimensional quantitative metrics (26). Herein, we integrated multi-parameter high-throughput feature mining with machine learning to precisely identify lesions.
In this study, we extracted 13 core radiomic features from contrast-enhanced T1-weighted MRI of 406 patients to develop a comprehensive model that integrates these features with clinical indicator data for the differentiation between cerebral tuberculomas and brain metastases of bronchial carcinoma. Notably, our model identified age, CEA levels, ring enhancement, and meningeal thickening on contrast-enhanced MRI as independent risk factors for distinguishing between these two conditions. The combined model demonstrated remarkable performance, achieving an AUC of 0.951 in the training cohort, 0.937 in the internal validation cohort, and 0.940 in the independent external validation cohort. Notably, IGRA was excluded from the combined model due to severe collinearity with Rad-score [variance inflation factor (VIF) >10]; this multicollinearity is attributable to shared pathological mechanisms. IGRA positivity reflects systemic T-cell secretion of IFN-γ in response to Mycobacterium tuberculosis infection; IFN-γ drives granuloma formation, inflammatory infiltration, necrosis, and tissue remodeling within cerebral tuberculomas (27). These morphological and microstructural alterations are captured and quantified by MRI radiomic features, which are aggregated into the Rad-score. Unlike tuberculomas, lung cancer brain metastases lack tuberculous granulomatous lesions and typically yield negative IGRA results. The overlapping immune-pathological pathways linking systemic IGRA signals and local imaging phenotypes result in predictive redundancy, thereby necessitating the exclusion of IGRA from the final combined model.
One prior study constructed a radiomic model using T1-weighted images of 97 tuberculomas and 78 metastases, proving that high-dimensional features can differentiate their pathological characteristics (28). Diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) analyses indicated markedly lower ADC values in the solid portions and walls of tuberculomas, reflecting disparities in cellular density and microarchitecture (29). Building on previous results, we adopted multi-parameter high-throughput feature mining and machine learning to better evaluate lesion heterogeneity, and established a clinical-radiomic fusion model using a larger sample size and clinical variables. By fully combining clinical indicators, morphological characteristics, texture features and high-dimensional hidden information, the model achieves precise lesion diagnosis via advanced algorithms, outperforming traditional diagnostic approaches based merely on morphology and signal changes.
To facilitate individualized classification and aid radiologists in distinguishing between cerebral tuberculomas and brain metastases of bronchial carcinoma, we also developed a bar graph visualization tool based on our combined model. Compared to conventional diagnostic methods, the classification predictions generated by this visualization tool are faster, more user-friendly, and more accurate, ultimately assisting radiologists in making precise diagnoses.
This study systematically assessed the sensitivity and specificity of the combined model through CV, thereby validating the model’s stability across diverse data distributions. Unlike traditional single-center or retrospective studies with small sample sizes, CV provides a more realistic reflection of the model’s performance in actual clinical scenarios. Existing literature indicates that radiomic models often exhibit decreased performance in independent external cohorts compared to their initial training sets, which can be attributed to factors such as sample distribution, imaging acquisition protocols, and patient heterogeneity (30,31). In our analysis, utilizing an external validation cohort, we found that the combined model achieved a sensitivity of 80.77% and specificity of 90.00%, outperforming the clinical model (sensitivity: 76.92%, specificity: 76.00%). Although the sensitivity of the combined model was lower than that of the radiomic model (92.31%), its specificity was superior (68.00%). The combined model’s specificity and PPV fell from 93.55% and 92.86% in internal validation to 90.00% and 80.77% in the external cohort. This discrepancy arises from differing cross-center distributions of clinical predictors. Radiomic features capture intrinsic lesion microstructural phenotypes and exhibit greater robustness across hospitals. By contrast, clinical indicators including serum CEA and meningeal thickening exhibit pronounced site-specific population heterogeneity. In the external cohort, several early-stage bronchial carcinoma brain metastases showed marginally elevated CEA and absent characteristic meningeal thickening; these lesions were accurately discriminated by radiomic signatures but misclassified upon incorporation of clinical variables, leading to reduced sensitivity. Nevertheless, clinical biomarkers effectively eliminated false-positive tuberculoma cases, which accounts for the higher specificity of the combined model in external validation.
This study elucidates the relationship between radiomic features and clinical molecular markers, particularly CEA. Our findings indicate a significant difference in CEA levels between tuberculomas and brain metastases of bronchial carcinoma (P<0.001), corroborating previously reported studies (32). Serum CEA was significantly elevated in brain metastases of bronchial carcinoma. In contrast, only mild CEA elevation was observed in patients with cerebral tuberculomas, and this elevation failed to reach statistical significance when compared with normal baseline values. This distinct intergroup disparity establishes CEA as a robust independent clinical predictor for differentiating the two intracranial lesions. Prior research has indicated that the serum CEA positivity rate is notably higher in patients with EGFR mutant lung adenocarcinoma compared to those with wild-type status, suggesting that CEA levels may serve as a useful biomarker for monitoring the efficacy of epidermal growth factor receptor tyrosine kinase inhibitor (EGFR-TKI) targeted therapies (33).
Emerging evidence indicates that tumor molecular subtypes can profoundly influence imaging phenotypes. For example, brain metastases from EGFR-mutant lung cancer typically present as well-defined, high-density lesions, whereas those harboring KRAS mutations exhibit greater imaging heterogeneity (32,33). Mechanistically, these differences may reflect underlying molecular pathways that regulate tumor proliferation, angiogenesis, and stromal remodeling—processes that shape imaging structure and texture. Despite such associations, significant phenotypic overlap across subtypes persists, highlighting the need for multi-omics integration to enhance diagnostic discrimination (23,34). Our systematic analysis reinforces the potential of radiomic features not only to differentiate brain lesion types but also to serve as non-invasive imaging biomarkers for molecular classification, offering promising avenues for targeted diagnostics and therapy.
Distinguishing multifocal intracranial lesions is clinically challenging. Although thoracic computed tomography (CT) provides valuable diagnostic information, it was excluded from our model. This study aimed to establish a brain MRI-only diagnostic tool, given that thoracic CT is not universally available at initial neurological evaluation. Additionally, inconsistent thoracic CT coverage and scanning timing in this retrospective cohort could have introduced unavoidable selection bias. Future studies may integrate thoracic CT data to further optimize model performance. Furthermore, shape-based radiomic features have shown incremental diagnostic potential in previous studies, a finding that merits further exploration in subsequent research (28).
This study has several limitations. First, our small single-center external cohort (comprising 76 patients with an unbalanced distribution of cases: 50 tuberculomas versus 26 metastases). Its small and imbalanced sample size may introduce sampling bias and reduce model generalizability. Second, the retrospective two-disease design carries inherent selection bias. Third, our analysis was restricted to contrast-enhanced T1-weighted, as standardized T2-weighted and DWI were unavailable across our retrospective cohort. Prior research has demonstrated that DWI distinguishes tuberculomas from small-cell lung cancer metastases but performs poorly for the non-small-cell lesions which constituted the majority of our cohort (35). T2-weighted reveals perilesional edema and intralesional necrosis, while DWI/ADC reflect cellular density and microstructural differences between the two lesions. The exclusion of these multi-parametric sequences limited the extraction of pathological features and the characterization of lesion heterogeneity, which likely contributed to the observed performance decline during external validation. Multi-modal radiomics integrating T1-weighted, T2-weighted and DWI can fuse complementary imaging signatures to generate more robust biomarkers, reduce sensitivity to inter-center scanning differences, and mitigate cross-cohort performance decay. We will acquire prospective multi-sequence data to optimize the model in future work. Fourth, only metastases originating from bronchial carcinoma were included; therefore, the findings cannot be generalized to intracranial metastases from other primary malignancies. Furthermore, our cohort did not include patients with concurrent cerebral metastases and tuberculomas. Although this combination is uncommon, the model’s diagnostic performance for such mixed lesions warrants further investigation. Lastly, dataset heterogeneity may compromise feature generalizability, and external validation was performed on a single cohort. Future work should incorporate larger, prospective cohorts, multi-modal imaging, and molecular data to further enhance model robustness and clinical utility.
In conclusion, we developed and validated a nomogram integrating clinical factors with contrast-enhanced MRI radiomic features, which accurately distinguishes cerebral tuberculomas from brain metastases of bronchial carcinoma—outperforming traditional clinical models. This non-invasive tool provides objective, quantitative decision support for early differentiation and individualized treatment planning.
Conclusions
In conclusion, based on previous studies, this study further suggests that a nomogram combining clinical factors and contrast-enhanced MRI radiomic features can effectively distinguish cerebral tuberculomas from brain metastases of bronchial carcinoma.
Supplementary
The article’s supplementary files as
Acknowledgments
The authors gratefully acknowledge DCPM (V5.49, Jingding Medical Technology Co., Ltd.) for technical support, and Dr. Shizhen Gou for technical support with the ImferScholar platform.
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. This retrospective study was approved by the Ethics Committees of Guangzhou Chest Hospital {No. Xiongyi Ethics [2023] No. 17} and Hunan Chest Hospital (No. LS2026011201), and individual consent for this retrospective analysis was waived.
Footnotes
Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1518/rc
Funding: This study was supported by the Guangzhou Science and Technology Program (grant Nos. 2025A03J3606, 2024A03J0583, 2024A03J0511), the Project of Guangdong Provincial Bureau of Traditional Chinese Medicine (grant No. 20251291), and the Scientific Research Program of the Beijing Medical Reward Foundation (grant No. YXJL-2024-0350-0245).
Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1518/coif). All authors report that this work was supported by the Guangzhou Science and Technology Program, Guangdong Provincial Bureau of Traditional Chinese Medicine and the Scientific Research Program of the Beijing Medical Reward Foundation. The authors have no other conflicts of interest to declare.
Data Sharing Statement
Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1518/dss
References
- 1. Khairy S, Alkhaibary A, AlQahtani S, Alkhani A. Cerebral tuberculoma. IDCases 2021;26:e01319.. 10.1016/j.idcr.2021.e01319 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Huntoon K, Musgrave N, Shaikhouni A, Elder J. Frequency of seizures in patients with metastatic brain tumors. Neurol Sci 2023;44:2501-7.. 10.1007/s10072-023-06695-y [DOI] [PubMed] [Google Scholar]
- 3. Lee RD, Atencio A. CNS tuberculomas. J Gen Intern Med 2023;38:2621.. 10.1007/s11606-023-08264-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.World Health Organization. Global tuberculosis report 2024. Geneva: World Health Organization; 2024. [accessed 24 August 2026]. Available online: https://www.who.int/teams/global-tuberculosis-programme/tb-reports/global-tuberculosis-report-2024.
- 5. Sidow NO. Large cerebral tuberculoma. Clin Case Rep 2024;12:e8827.. 10.1002/ccr3.8827 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Siegel RL, Kratzer TB, Giaquinto AN, Sung H, Jemal A. Cancer statistics, 2025. CA Cancer J Clin 2025;75:10-45.. 10.3322/caac.21871 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Mackel CE, Rosenberg H, Varma H, Uhlmann EJ, Vega RA, Alterman RL. Intracranial Metastasis of Extracranial Chondrosarcoma: Systematic Review With Illustrative Case. Brain Tumor Res Treat 2023;11:103-13.. 10.14791/btrt.2023.0003 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. Papadimitriou K, Kiss-Bodolay D, Hedjoudje A, Millan DS, Simonin A, Fournier JY, Huscher K. Late metachronous cerebral metastasis of pancreatic adenocarcinoma of the tail of the pancreas: a case report. J Med Case Rep 2022;16:144.. 10.1186/s13256-022-03314-w [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Müller SJ, Khadhraoui E, Ernst M, Rohde V, Schatlo B, Malinova V. Differentiation of multiple brain metastases and glioblastoma with multiple foci using MRI criteria. BMC Med Imaging 2024;24:3.. 10.1186/s12880-023-01183-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Müller SJ, Khadhraoui E, Henkes H, Ernst M, Rohde V, Schatlo B, Malinova V. Differentiation between multifocal CNS lymphoma and glioblastoma based on MRI criteria. Discov Oncol 2024;15:397.. 10.1007/s12672-024-01266-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Mertiri L, Freiling JT, Desai NK, Kralik SF, Huisman TAGM. Pediatric and adult meningeal, parenchymal, and spinal tuberculosis: A neuroimaging review. J Neuroimaging 2024;34:179-94.. 10.1111/jon.13177 [DOI] [PubMed] [Google Scholar]
- 12. Chinthala AS, Obeng-Gyasi B, Virgin KL, Deckert M, Mao G. Brain metastasis mimicking brain abscess: illustrative case and systematic review. J Neurosurg Case Lessons 2025;10:CASE25528.. 10.3171/CASE25528 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Garg R, Shukla S, Garg RK. Challenges and advancements in the diagnosis of central nervous system tuberculosis. Indian J Tuberc 2026;73:270-6.. 10.1016/j.ijtb.2025.08.011 [DOI] [PubMed] [Google Scholar]
- 14. Ma Q, Yi Y, Liu T, Wen X, Shan F, Feng F, Yan Q, Shen J, Yang G, Shi Y. MRI-based radiomics signature for identification of invisible basal cisterns changes in tuberculous meningitis: a preliminary multicenter study. Eur Radiol 2022;32:8659-69.. 10.1007/s00330-022-08911-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Xiong Z, Qiu J, Liang Q, Jiang J, Zhao K, Chang H, Lv C, Zhang W, Li B, Ye J, Li S, Peng S, Sun C, Chen S, Long D, Shu X. Deep learning models for rapid discrimination of high-grade gliomas from solitary brain metastases using multi-plane T1-weighted contrast-enhanced (T1CE) images. Quant Imaging Med Surg 2024;14:5762-73.. 10.21037/qims-24-380 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Demircioğlu A. Benchmarking Feature Selection Methods in Radiomics. Invest Radiol 2022;57:433-43.. 10.1097/RLI.0000000000000855 [DOI] [PubMed] [Google Scholar]
- 17. Demircioğlu A. Benchmarking feature projection methods in radiomics. Sci Rep 2025;15:32368.. 10.1038/s41598-025-16070-w [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Fellner A, Gavriel H, Pitaro J, Muallem Kalmovich L. Clinical parameters predicting tonsillar malignancy. Eur Arch Otorhinolaryngol 2020;277:1779-83.. 10.1007/s00405-020-05873-4 [DOI] [PubMed] [Google Scholar]
- 19. Zhu D, Zhang M, Li Q, Liu J, Zhuang Y, Chen Q, Chen C, Xiang Y, Zhang Y, Yang Y. Can perihaematomal radiomics features predict haematoma expansion?. Clin Radiol 2021;76:629. [DOI] [PubMed] [Google Scholar]
- 20. Izabella Abreu de Melo M, Rodrigues Correa C, da Silva Cunha P, Miranda de Góes A, Assis Gomes D, Silva Ribeiro de Andrade A. DNA aptamers selection for carcinoembryonic antigen (CEA). Bioorg Med Chem Lett 2020;30:127278.. 10.1016/j.bmcl.2020.127278 [DOI] [PubMed] [Google Scholar]
- 21. Ren C, Tang J, Xia L. Interferon gamma release assays for diagnosis of osteoarticular tuberculosis: A systematic review and meta-analysis. PLoS One 2022;17:e0269234.. 10.1371/journal.pone.0269234 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Desai N, Krishnan R, Rukmangadachar L. Central Nervous System Tuberculosis Presenting With Multiple Ring-Enhancing Lesions: A Diagnostic Challenge. Cureus 2022;14:e21819.. 10.7759/cureus.21819 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. de Medeiros FC, de Albuquerque LA, Pittella JE, de Souza RB, Gomes Neto AP, Christo PP. Open-ring enhancement in pseudotumoral multiple sclerosis: important radiological aspect. Case Rep Neurol Med 2014;2014:951690.. 10.1155/2014/951690 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Bravo-Tsri AEB, Konaté I, Kouassi KPB, Acko-Ohui EV, Goulé-Bi AR, Isart D, Tanoh EK, Vangah MK, Kouadio FA, Yao BL. Meningeal tuberculoma mimicking a brain tumor. Radiol Case Rep 2021;16:284-8. Erratum in: Radiol Case Rep. 2023 Jan 25;18(4):1643-1644.. 10.1016/j.radcr.2020.11.028 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. McMahon P, Pisapia DJ, Schweitzer AD, Heier L, Souweidane MM, Roytman M. Central nervous system tuberculoma mimicking a brain tumor: A case report. Radiol Case Rep 2024;19:414-7.. 10.1016/j.radcr.2023.10.042 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Abdel Razek AAK, Alksas A, Shehata M, AbdelKhalek A, Abdel Baky K, El-Baz A, Helmy E. Clinical applications of artificial intelligence and radiomics in neuro-oncology imaging. Insights Imaging 2021;12:152.. 10.1186/s13244-021-01102-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Yang Y, Wang HJ, Hu WL, Bai GN, Hua CZ. Diagnostic Value of Interferon-Gamma Release Assays for Tuberculosis in the Immunocompromised Population. Diagnostics (Basel) 2022;12:453.. 10.3390/diagnostics12020453 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Wang A, Qiu X, Chen Q, Qiu Y, Wang Y, Hu Y, Wang J, Zhan M, Zhu H. MRI quantitative imaging biomarkers in differentiating brain parenchymal tuberculoma and lung cancer brain metastases. Eur J Med Res 2025;30:1239.. 10.1186/s40001-025-03476-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Cui L, Qin Z, Sun S, Feng W, Hou M, Yu D. Diffusion-weighted imaging-based radiomics model using automatic machine learning to differentiate cerebral cystic metastases from brain abscesses. J Cancer Res Clin Oncol 2024;150:132.. 10.1007/s00432-024-05642-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Cuocolo R, Stanzione A, Faletti R, Gatti M, Calleris G, Fornari A, Gentile F, Motta A, Dell'Aversana S, Creta M, Longo N, Gontero P, Cirillo S, Fonio P, Imbriaco M. MRI index lesion radiomics and machine learning for detection of extraprostatic extension of disease: a multicenter study. Eur Radiol 2021;31:7575-83.. 10.1007/s00330-021-07856-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Koska IO, Ozcan HN, Tan AA, Beydogan B, Ozer G, Oguz B, Haliloglu M. Radiomics in differential diagnosis of Wilms tumor and neuroblastoma with adrenal location in children. Eur Radiol 2024;34:5016-27.. 10.1007/s00330-024-10589-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Di Stefano AL, Picca A, Saragoussi E, Bielle F, Ducray F, Villa C, et al. Clinical, molecular, and radiomic profile of gliomas with FGFR3-TACC3 fusions. Neuro Oncol 2020;22:1614-24.. 10.1093/neuonc/noaa121 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33. Xu J, Yang Y, Gao Z, Song T, Ma Y, Yu X, Shi C. Distinguishing EGFR mutation molecular subtypes based on MRI radiomics features of lung adenocarcinoma brain metastases. Clin Neurol Neurosurg 2024;240:108258.. 10.1016/j.clineuro.2024.108258 [DOI] [PubMed] [Google Scholar]
- 34. Liu Q, Hu P. Radiogenomic association of deep MR imaging features with genomic profiles and clinical characteristics in breast cancer. Biomark Res 2023;11:9.. 10.1186/s40364-023-00455-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35. Müller SJ, Khadhraoui E, Neef NE, Riedel CH, Ernst M. Differentiation of brain metastases from small and non-small lung cancers using apparent diffusion coefficient (ADC) maps. BMC Med Imaging 2021;21:70.. 10.1186/s12880-021-00602-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
The article’s supplementary files as
Data Availability Statement
Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1518/dss