IPBronch.com | Interventional Pulmonology

🎯 Background & Rationale

Differentiating between pulmonary sarcoidosis and mediastinal lymph node tuberculosis (TB) remains a frequent diagnostic dilemma in the bronchoscopy suite, particularly in geographic regions where TB is endemic. Both granulomatous diseases present with remarkably similar clinical features, radiographic patterns, and cytomorphological appearances on EBUS-TBNA samples. This publication addresses the clinical gap of whether advanced imaging post-processing—specifically, endobronchial ultrasound (EBUS)-derived radiomics combined with machine learning—can reliably distinguish between these two benign granulomatous etiologies without waiting for delayed microbiological cultures.

👥 Study Design & Population

Study Type: Retrospective Cohort Study
Population: A total of 170 patients (100 diagnosed with pulmonary sarcoidosis and 70 diagnosed with mediastinal lymph node tuberculosis) who underwent EBUS evaluation and subsequent definitive diagnosis.
Intervention/Comparison: Extraction of radiomics features from EBUS images combined with clinical characteristics, analyzed via a support vector machine (SVM) algorithm to differentiate sarcoidosis from lymph node TB. The cohort was randomly split into a training group (n = 119) and a validation group (n = 51).

📈 Methodology & Rigor

The authors utilized least absolute shrinkage and selection operator (LASSO) regression to isolate 7 stable radiomics features with nonzero coefficients from EBUS ultrasound images, which were combined with 4 statistically significant clinical features. A support vector machine (SVM) algorithm was trained to construct the diagnostic model. Diagnostic performance was evaluated using receiver operating characteristic (ROC) curves, area under the curve (AUC), accuracy, sensitivity, and specificity in both training and validation cohorts.

🔬 Key Findings [or Planned Endpoints]

The EBUS-based radiomics and clinical SVM model demonstrated robust performance:

Training Group: Achieved a ROC AUC of 0.909 (95% CI: 0.897–0.922), with an overall accuracy of 88.2%, sensitivity of 82.4%, and specificity of 92.6%.
Validation Group: Maintained high performance with a ROC AUC of 0.917 (95% CI: 0.901–0.934), yielding an accuracy of 80.4%, sensitivity of 68.4%, and specificity of 87.5%.
Inputs: The final model relied on a combination of 7 radiomics features and 4 clinical variables.

⚖️ Critical Appraisal

While the statistical performance metrics (AUC > 0.90) are promising, several limitations warrant consideration. First, the retrospective single-center design introduces potential selection bias and limits the generalizability of the model to other ultrasound platforms or operator techniques. EBUS image acquisition can be inherently subjective and variable depending on gain settings, pressure applied to the bronchial wall, and transducer frequency. Furthermore, external validation across multi-center cohorts using diverse EBUS hardware is essential before widespread clinical adoption can be recommended.

💡 The Clinical Bottom Line

This study highlights the burgeoning role of artificial intelligence and radiomics in interventional pulmonology. While EBUS-TBNA remains the gold standard for tissue acquisition, an EBUS-derived radiomics model could eventually serve as a real-time auxiliary decision-support tool during bronchoscopy—potentially offering immediate guidance on differential diagnoses while awaiting time-consuming TB cultures or molecular assays. However, this remains a preliminary tool and should not yet replace traditional cytopathological and microbiological confirmation.

Introduction: To explore the feasibility of an ultrasound radiomics machine learning model based on endobronchial ultrasound (EBUS) for differentiating pulmonary sarcoidosis from mediastinal lymph node tuberculosis. Methods: Clinical characteristics and ultrasound image data from 100 patients diagnosed with pulmonary sarcoidosis and 70 patients diagnosed with mediastinal lymph node tuberculosis were collected. Statistical analysis was performed to compare clinical features, such as age, gender, smoking history, and lymph node size and so on, between the two groups. The least absolute shrinkage and selection operator (Lasso) was used to analyze the radiomics features extracted from EBUS-based ultrasound images. A support vector machine (SVM) algorithm was applied to establish an EBUS-based radiomics model, and clinical features with statistically significant were incorporated to optimize the model. A total of 170 lymph nodes were randomly divided into training group (n=119) and validation group (n=51), with the diagnostic performance of the model assessed using receiver operating characteristic (ROC) curves and the area under the curve (AUC), accuracy, sensitivity, and specificity. Results: Seven stable radiomics features with non-zero coefficients and four clinical features were selected as inputs for the model. The SVM model demonstrated great performance in both groups. In the training group, the ROC AUC of the SVM model was 0.909 (95% CI: 0.897-0.922), with an accuracy of 88.2%, sensitivity of 82.4%, and specificity of 92.6%. In the validation group, the ROC AUC was 0.917 (95% CI: 0.901-0.934), with an accuracy of 80.4%, sensitivity of 68.4%, and specificity of 87.5%. Conclusion: The SVM model based on EBUS ultrasound radiomics and clinical data shows promising feasibility for differentiating pulmonary sarcoidosis from mediastinal lymph node tuberculosis. It exhibits significant potential in clinical practice and provides a new method for the early diagnosis of these conditions.