RepAir: A Framework for Airway Segmentation and Discontinuity Correction in CT

Accurate airway segmentation from chest computed tomography (CT) scans is essential for quantitative lung analysis, yet manual annotation is impractical and many automated U-Net-based methods yield disconnected components that hinder reliable biomarker extraction. We present RepAir, a three-stage framework for robust 3D airway segmentation that combines an nnU-Net-based network with anatomically informed topology correction. The segmentation network produces an initial airway mask, after which a skeleton-based algorithm identifies potential discontinuities and proposes reconnections. A 1D convolutional classifier then determines which candidate links correspond to true anatomical branches versus false or obstructed paths. We evaluate RepAir on two distinct datasets: ATM'22, comprising annotated CT scans from predominantly healthy subjects and AeroPath, encompassing annotated scans with severe airway pathology. Across both datasets, RepAir outperforms existing 3D U-Net-based approaches such as Bronchinet and NaviAirway on both voxel-level and topological metrics, and produces more complete and anatomically consistent airway trees while maintaining high segmentation accuracy.

💡 Research Summary

The paper introduces RepAir, a three‑stage framework designed to produce anatomically consistent airway segmentations from chest CT scans. The authors identify a persistent problem in existing deep‑learning airway segmentation methods: while 3D U‑Net variants such as Bronchinet and NaviAirway achieve high voxel‑wise overlap, they frequently generate fragmented airway trees, especially in scans with severe disease where small, high‑generation bronchi are hard to detect. RepAir addresses this by coupling a robust nnU‑Net segmentation backbone with a topology‑aware correction pipeline.

Stage 1 employs nnU‑Net, trained on 3‑D patches of size 128 × 160 × 112, using a combined Dice + cross‑entropy loss and extensive data augmentation (rotations, scaling, noise, blur, intensity shifts, mirroring). This backbone provides strong feature extraction across varying resolutions and maintains high voxel‑level accuracy.

Stage 2 extracts a skeleton from the initial airway mask and searches for discontinuities. For each skeleton endpoint, the algorithm scans a predefined radius for nearby skeleton voxels that could serve as connection partners. Candidate links are ranked by distance (shorter preferred), angular deviation (sharp, implausible bends penalized), and a cycle‑avoidance rule that preserves the tree structure. The selected links are dilated to match the local airway diameter, producing anatomically plausible candidate reconnections.

Stage 3 uses a 1‑D convolutional neural network to decide whether each candidate link corresponds to a true airway branch, lung parenchyma, or an obstructed airway. Training data consist of 20 000 patches per class: true airway samples are taken from ground‑truth centerlines, parenchyma samples from non‑airway lung tissue, and obstruction samples are generated by adding high‑density noise to healthy airway segments. The 1‑D CNN receives ordered voxel intensities along the candidate centerline and its neighboring segments, learning to discriminate based on intensity patterns and geometric continuity. Only links classified as true airway are merged into the final mask, preventing false connections and preserving anatomical realism.

The authors evaluate RepAir on two publicly available datasets. ATM’22 contains 279 inspiratory CT scans (209 training, 70 testing) mainly from healthy subjects and a few COVID‑19 cases, with expert airway annotations. AeroPath comprises 27 scans from patients with severe pathologies (malignant tumors, sarcoidosis, emphysema) and similarly detailed annotations. Both datasets present different challenges: ATM’22 tests performance on typical anatomy, while AeroPath stresses robustness to disease‑induced distortions.

Performance is measured using Dice coefficient, true positive rate (TPR), false positive rate (FPR), Jaccard index (JI), Hausdorff distance (HD), and tree‑detected rate (TD), the latter quantifying the proportion of ground‑truth airway centerline length recovered. On ATM’22, RepAir achieves Dice 0.93 ± 0.02, JI 0.88 ± 0.04, TD 0.92 ± 0.06, and HD 42.75 ± 20.53 mm, surpassing Bronchinet (Dice 0.92, JI 0.85, TD 0.82, HD 44.61) and NaviAirway (Dice 0.80, JI 0.67, TD 0.47, HD 69.96). RepAir also records the lowest FPR (0.55 × 10⁻⁴), indicating fewer spurious detections. Visual inspection shows markedly fewer false‑negative regions, especially in peripheral bronchi.

On the more challenging AeroPath dataset, RepAir maintains superior results (Dice 0.88 ± 0.03, JI 0.78 ± 0.05, TD 0.90 ± 0.07, HD 50.15 ± 17.22 mm) compared with Bronchinet (Dice 0.62, TD 0.25, HD 173.92) and NaviAirway (Dice 0.84, TD 0.63, HD 79.86). The method’s ability to recover high‑generation airways despite severe obstruction demonstrates strong generalization from the ATM’22 training set without fine‑tuning.

Runtime analysis shows RepAir’s inference time is about 3 minutes per scan, slightly slower than Bronchinet due to the deeper nnU‑Net encoder‑decoder, but marginally faster than NaviAirway (by ~3 seconds). This modest overhead is justified by the substantial gains in topological completeness.

Key contributions of the work include: (1) integrating a state‑of‑the‑art nnU‑Net segmentation model with a dedicated topology‑correction stage, (2) designing a skeleton‑based discontinuity detector that respects anatomical constraints, (3) employing a lightweight 1‑D CNN to validate candidate connections, thereby eliminating false links while restoring missing branches, and (4) demonstrating consistent improvements across both healthy and diseased cohorts, indicating broad applicability. The authors note that the correction module is agnostic to the underlying segmentation network, allowing RepAir to be combined with any airway predictor.

Future directions suggested involve embedding connectivity constraints directly into the loss function for end‑to‑end training, extending the framework to multi‑task learning that simultaneously segments airway walls, and evaluating the impact of more complete airway trees on longitudinal disease monitoring and therapeutic response assessment. In summary, RepAir provides a practical, high‑performance solution for automated airway segmentation, bridging the gap between voxel‑wise accuracy and topological fidelity, and holds promise for advancing quantitative lung imaging in both research and clinical settings.