Artificial Intelligence and Endobronchial Ultrasound for Lymph-Node Detection and Characterization: A Pioneer Multicenter Transatlantic Study.
Authors
Affiliations (14)
Affiliations (14)
- Department of Community Medicine, Information and Health Decision Sciences, Faculty of Medicine, University of Porto, 4200-319 Porto, Portugal.
- Precision Medicine Unit, Department of Gastroenterology, Unidade Local de Saúde de São João, 4200-319 Porto, Portugal.
- World Gastroenterology Organization Gastroenterology and Hepatology Training Center, 4200-319 Porto, Portugal.
- Department of Pneumology, Hospital Universitario Puerta de Hierro Majadahonda, 28222 Madrid, Spain.
- Department of Thoracic Surgery, Instituto Português de Oncologia-Porto, 4200-072 Porto, Portugal.
- Gastrointestinal Endoscopy Unit, Department of Gastroenterology, Hospital das Clínicas da Faculdade de Medicina da Universidade de São Paulo, São Paulo 05403-000, Brazil.
- Department of Thoracic Surgery, Hospital São Luiz-Rede D'Or, São Paulo 04544-000, Brazil.
- Department of Medicine, Faculty of Medicine, University of Porto, 4200-319 Porto, Portugal.
- Department of Digestive Endoscopy, Hospital Universitario Puerta de Hierro Majadahonda, 28222 Madrid, Spain.
- Department of Mechanical Engineering, Faculty of Engineering, University of Porto, 4200-465 Porto, Portugal.
- Department of Pneumology, Hospital São Luiz-Rede D'Or, São Paulo 04544-000, Brazil.
- Department of Pneumology, Hospital das Clínicas da Faculdade de Medicina da Universidade de São Paulo, São Paulo 05403-900, Brazil.
- Department of Surgery and Physiology, Faculty of Medicine, University of Porto, 4200-319 Porto, Portugal.
- Cardiothoracic Surgery, Unidade Local de Saúde de São João, 4200-437 Porto, Portugal.
Abstract
<b>Background/Objectives</b>: Accurate mediastinal and hilar lymph node (LN) assessment is central to lung cancer staging and treatment selection. Endobronchial ultrasound-guided transbronchial needle aspiration (EBUS-TBNA) has become the standard minimally invasive procedure for mediastinal staging, enabling real-time identification, characterization and sampling of LNs. Although effective, this technique remains operator dependent, and subject to interobserver variability. This study aims to report the development of a Vision Transformer-based AI model for automatic LN detection and classification on EBUS images. <b>Methods</b>: A multicenter, multidevice, retrospective study, using EBUS images of three referral centers (Spain and Brazil). The dataset included images of patients with benign and malignant LNs. Bounding-box annotations were applied for LN localization, and LNs were classified as benign or malignant according to histopathologic reference standards. Detection performance was evaluated with mean average precision, at an intersection over union threshold of 0.5 (mAP50). Characterization performance was assessed using accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC). <b>Results</b>: The dataset included 6492 images from 27 patients, of whom 16 had malignant LNs. In the test set, the model achieved a mAP50 of 0.79 for LN detection. For malignant characterization, the model achieved an accuracy of 90.6%, sensitivity of 85.4%, and an AUROC of 0.938. <b>Conclusions</b>: The results of this study support the feasibility of AI-assisted EBUS interpretation for nodal staging workflows and raise the possibility that AI-assisted EBUS could enhance diagnostic performance, while reducing operator variability. Larger prospective studies are needed before clinical implementation.