Clinical Need Airway nodules -- growths arising within or adjacent to the bronchial tubes -- are a distinct category of lung lesion that are easy to miss on CT scans. Unlike parenchymal lung nodules, they require different clinical management and can signal primary endobronchial tumors or metastases.
Study Design This study developed and evaluated a deep learning-based computer-aided detection (DL-CAD) system for automatically identifying airway nodules in chest CT scans. The dataset included 320 patients: 160 with confirmed airway nodules and 160 controls without airway nodules.
AI Architecture The DL-CAD pipeline combined YOLOv8 -- a state-of-the-art real-time object detection network -- with ConvNeXt, a modern convolutional architecture, to perform two-stage detection and classification of airway nodule candidates.
Key Results The system achieved an overall sensitivity of 75.1% across all nodule types, with sensitivity of 79% specifically for tumors at a false positive rate of 0.25 false positives per scan -- a threshold considered acceptable for clinical screening tools.
Stage 1 -- Candidate Detection YOLOv8 (You Only Look Once, version 8) processed chest CT slices to generate initial candidate regions potentially containing airway nodules. This fast, anchor-free detector was chosen for its ability to handle small objects in complex backgrounds.
Stage 2 -- False Positive Reduction ConvNeXt, a pure convolutional network designed to rival transformer architectures in accuracy, served as a classifier to evaluate each YOLOv8 candidate and reject false positives while retaining true airway nodule detections.
Airway Segmentation Before detection, an automated airway tree segmentation algorithm extracted the bronchial tree from each CT scan. Candidate nodule regions were then restricted to the vicinity of airways, reducing the search space and false positive burden.
3D Context Integration The pipeline processed multiple consecutive CT slices to capture three-dimensional context for each candidate region. This multi-slice approach improved detection of small nodules that may only appear on a few slices.
FROC Methodology Performance was assessed using Free-Response Receiver Operating Characteristic (FROC) analysis, which plots sensitivity against the average number of false positives per scan. This is the standard evaluation method for nodule detection systems.
Sensitivity by Nodule Type The system performed best for carcinoid tumors and primary endobronchial malignancies, with lower sensitivity for inflammatory lesions and mucus plugging that can mimic airway nodules on CT.
Comparison to Radiologists At the 0.25 false positives per scan operating point, the AI system achieved sensitivity comparable to experienced chest radiologists, suggesting clinical utility as an assistive tool in reading chest CT studies.
Nodule Size Impact Performance improved with increasing nodule size. Nodules smaller than 5 mm were the most challenging to detect reliably, which is consistent with the limits of CT resolution and reader performance for small airway lesions.
Case Selection Cases were selected from a tertiary hospital's database with pathological or bronchoscopic confirmation of airway nodule status. Controls were matched on age, sex, and CT acquisition parameters to reduce confounding.
Annotation Process Each airway nodule was annotated by two chest radiologists with a third senior radiologist resolving disagreements. Both bounding box locations and nodule classification (tumor, inflammatory, other) were recorded.
Data Augmentation Because airway nodules are rare findings, extensive data augmentation was applied during training including rotation, flipping, intensity variation, and synthetic nodule insertion to address class imbalance.
Training-Validation Split The dataset was divided into training (70%), validation (15%), and test (15%) sets, with the test set withheld until final model evaluation to provide an unbiased estimate of real-world performance.
Reducing Missed Diagnoses Airway nodules are frequently missed on CT because radiologists typically focus on lung parenchyma. An automated alert system drawing attention to the airway tree could reduce these missed diagnoses.
Bronchoscopy Guidance When the AI system flags an airway nodule, the output includes location coordinates that could guide bronchoscopy planning -- the procedure used to biopsy these lesions under direct visualization.
Screening Applications As low-dose CT lung cancer screening programs expand, adding airway nodule detection to screening protocols could identify endobronchial tumors that would otherwise be missed in standard screening reads.
Workflow Efficiency The system processes a full chest CT scan in seconds, making it feasible to apply to all chest CT studies as a background quality check without adding significant burden to radiology workflows.
Single-Center Data The dataset came from a single institution, limiting generalizability to different CT scanners, acquisition protocols, and patient populations. Multi-center external validation is the critical next step.
Class Imbalance Challenges Even within the airway nodule group, some subtypes (such as rare benign tumors) were represented by very few cases. Performance on rare nodule types may not reflect real-world accuracy.
Sensitivity Ceiling at Small Sizes For nodules below 5 mm, detection sensitivity remains substantially lower than for larger nodules. Improving small nodule detection would require higher resolution CT inputs or specialized architectures.
Prospective Clinical Trial A randomized trial comparing radiologist performance with and without AI assistance would definitively establish the clinical value of the system in terms of patient outcomes such as time to diagnosis and treatment.