Artificial intelligence guided automated image analysis of endobronchial lesions during white-light bronchoscopy to predict moderate to severe airway bleeding
Related Experiment Video
Updated: Nov 3, 2026
Robotic-assisted Bronchoscopy Combined with Multimodal Imaging for Targeted Lung Cryobiopsies
Published on: July 19, 2024
Saikat Banerjee1, Krishnapriya S Kumar2, Ruchi Dua3
1Consultant Pulmonologist, Insitute of Pulmocare and Research, Kolkata, West Bengal, India.
Background:
Bleeding during flexible bronchoscopy is an important procedural complication, particularly in patients with endobronchial lesions. Early identification of lesions at higher risk may assist bronchoscopists in procedural planning and preparedness. Artificial intelligence (AI)-based image analysis may enable real-time risk stratification during bronchoscopy.
Materials And Methods:
In this exploratory observational study, 250 patients with endobronchial lesions undergoing flexible bronchoscopy were analysed. Clinical characteristics, symptoms, bronchoscopic lesion features, and radiological findings were recorded. Procedural bleeding was graded and categorized as nil/mild or moderate/severe. Predictors were identified using Chi-square and multivariate logistic regression analyses. Machine-learning models (Random Forest (RF), XGBoost, CatBoost, and an ensemble classifier) were developed using clinical and bronchoscopic features. Model performance was assessed using accuracy, precision, recall, and area under the receiver operating characteristic curve (AUC).
Results:
Of 250 patients, 82.8% were male and 60.4% were smokers. Mild, moderate, and severe bleeding occurred in 58.0%, 37.2%, and 3.2% of cases, respectively. Moderate haemoptysis prior to bronchoscopy ( P = 0.028), upper-lobe lesion location on bronchoscopy ( P = 0.013) and computed tomography (CT) ( P = 0.027), and vascular lesion appearance ( P = 0.046) were independently associated with higher bleeding risk. The ensemble model demonstrated the best performance (accuracy 85.0%, AUC 0.88), with higher recall (0.76) compared with individual models.
Conclusion:
Clinical and bronchoscopic features can help identify lesions at higher risk of bleeding during bronchoscopy. AI-based models showed promising performance and may serve as real-time intra-procedural decision-support tools. Further multicentre studies with external validation are required.
