AirwayNet: A Voxel-Connectivity Aware Approach for Accurate Airway Segmentation Using Convolutional Neural Networks

Guang-Zhong Yang; Hao Zheng; Jie Yang; Mali Shen; Mingjian Chen; Xiaolin Huang; Yue-Min Zhu; Yulei Qin; Yun Gu

arxiv: 1907.06852 · v1 · pith:LPLGE2LBnew · submitted 2019-07-16 · 📡 eess.IV · cs.CV

AirwayNet: A Voxel-Connectivity Aware Approach for Accurate Airway Segmentation Using Convolutional Neural Networks

Yulei Qin , Mingjian Chen , Hao Zheng , Yun Gu , Mali Shen , Jie Yang , Xiaolin Huang , Yue-Min Zhu

show 1 more author

Guang-Zhong Yang

This is my paper

Pith reviewed 2026-05-24 20:54 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords airway segmentationconvolutional neural networksvoxel connectivityCT scanslung imagingmedical image segmentationtree-like structures

0 comments

The pith

AirwayNet improves CT airway segmentation by predicting how each voxel connects to its 26 neighbors instead of simple binary labeling.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces AirwayNet to segment airways in lung CT scans more accurately. Standard CNNs struggle with the tree-like branching and connections in airways. By turning the task into 26 separate predictions of voxel connectivity, the network learns both the structure and how neighboring voxels relate. Adding lung distance maps and voxel coordinates provides extra context. This leads to better performance than previous methods on airway extraction tasks.

Core claim

By converting the conventional binary segmentation task into 26 tasks of connectivity prediction, AirwayNet learns both airway structure and the relationship between neighboring voxels, achieving superior segmentation performance compared to existing approaches.

What carries the argument

Voxel-connectivity aware CNN that models 26-neighbor connectivity predictions, augmented with lung distance map and voxel coordinates as additional inputs.

If this is right

Airway segmentation becomes more accurate for pulmonary disease diagnosis and endobronchial navigation.
The network better captures the tree-like pattern of airways through explicit connectivity modeling.
Context knowledge from distance maps and coordinates enhances semantic understanding in the segmentation.
Superior performance demonstrates the value of awareness of voxel connectivity over standard binary segmentation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If connectivity modeling works here, similar transformations could apply to other tubular structures like blood vessels in medical imaging.
Testing on diverse CT datasets with varying resolutions might reveal how robust the 26-connectivity approach is across different scanners.
The method could reduce manual effort in extracting airways for clinical use if integrated into diagnostic software.

Load-bearing premise

Converting the segmentation task into 26 connectivity prediction tasks will enable the network to perceive the tree-like pattern and comprehend airway connectivity better than standard binary segmentation.

What would settle it

A direct comparison on standard airway segmentation benchmarks where AirwayNet does not outperform existing CNN methods would disprove the claim that connectivity awareness is the key improvement.

Figures

Figures reproduced from arXiv: 1907.06852 by Guang-Zhong Yang, Hao Zheng, Jie Yang, Mali Shen, Mingjian Chen, Xiaolin Huang, Yue-Min Zhu, Yulei Qin, Yun Gu.

**Figure 1.** Figure 1: Flowchart of the proposed AirwayNet. 2.1 Connectivity Modeling Using Binary Ground-truth Labels In a three-dimensional (3-D) CT, 26-connectivity describes well the relation between one voxel and its 26 neighbors (see [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Illustration of 26-connectivity modeling. The binary ground-truth of airway (Dim: 1×Z ×H ×W) is transformed into a connectivity label (Dim: 26×Z ×H ×W). 2.2 CT Volume Pre-processing One challenge in airway segmentation is that the foreground voxels only occupy a small proportion of all CT voxels. To avoid feature learning from irrelevant parts (e.g., ribs and skin), we restrict the valid airway candidate r… view at source ↗

**Figure 3.** Figure 3: The architecture of the 3-D CNNs used in the proposed AirwayNet. The number of channels is denoted above each feature map [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Comparison of airway segmentation results. Local bronchial branches are highlighted with circles and zoomed in to better see performance difference. to the introduction of connectivity modeling and semantic information, making the model more sensitive and perceptive to airway than background. Meanwhile, the improvement in DSC and TPR proves that it is worthwhile to enable the model being aware of connecti… view at source ↗

read the original abstract

Airway segmentation on CT scans is critical for pulmonary disease diagnosis and endobronchial navigation. Manual extraction of airway requires strenuous efforts due to the complicated structure and various appearance of airway. For automatic airway extraction, convolutional neural networks (CNNs) based methods have recently become the state-of-the-art approach. However, there still remains a challenge for CNNs to perceive the tree-like pattern and comprehend the connectivity of airway. To address this, we propose a voxel-connectivity aware approach named AirwayNet for accurate airway segmentation. By connectivity modeling, conventional binary segmentation task is transformed into 26 tasks of connectivity prediction. Thus, our AirwayNet learns both airway structure and relationship between neighboring voxels. To take advantage of context knowledge, lung distance map and voxel coordinates are fed into AirwayNet as additional semantic information. Compared to existing approaches, AirwayNet achieved superior performance, demonstrating the effectiveness of the network's awareness of voxel connectivity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

AirwayNet recasts binary airway segmentation as 26 connectivity predictions plus extra maps, but the abstract supplies no ablations or results to show the connectivity change is what helps.

read the letter

The main thing to know is that this paper changes the output of a segmentation network from one binary airway mask to 26 separate connectivity predictions per voxel, while also feeding in a lung distance map and voxel coordinates as extra channels. The goal is to make the model explicitly learn neighbor relations in the airway tree instead of relying on standard segmentation to preserve connectivity on its own. That reformulation is the concrete technical step they take. It is a reasonable response to the well-known problem that thin, branching structures like airways are easy for CNNs to break or miss. The description in the abstract is straightforward and avoids any circular definitions or invented quantities. The approach engages directly with the topology issue rather than treating it as a generic segmentation task. The soft spot is the missing evidence for the central assumption. The abstract states that the connectivity modeling lets the network perceive tree-like patterns better, yet it introduces the 26-task head, the distance map, and the coordinate channels all at once. No ablation is described that keeps the architecture, training, and extra inputs fixed while switching only the output and loss back to ordinary binary segmentation. Without that isolation, any performance gain could come from the added semantic channels rather than the connectivity formulation itself. The abstract also gives no datasets, metrics, baselines, or numbers, so the claim of superior performance cannot be checked. If the full paper contains those ablations and quantitative comparisons, the contribution would be easier to assess. This work is aimed at researchers doing airway or vessel segmentation on CT for pulmonary applications. A reader already working on that narrow clinical task might find the connectivity output idea worth testing, but the paper is too specialized and under-specified for wider use. I would not bring it to a reading group on the current information. I would not cite it. It deserves peer review if the full version includes proper experiments and ablations, because the underlying idea is clear and addresses a real limitation even if the current write-up leaves the key attribution untested.

Referee Report

2 major / 1 minor

Summary. The paper proposes AirwayNet, a CNN-based method for airway segmentation from CT scans. It converts the conventional binary segmentation task into 26 per-voxel connectivity prediction tasks to enable the network to better perceive tree-like airway structures and voxel connectivity. Lung distance maps and voxel coordinates are added as extra input channels for context. The authors report superior performance over existing approaches, attributing gains to the connectivity-aware formulation.

Significance. If the performance gains are reproducible and attributable to the connectivity modeling, the approach could meaningfully improve automatic extraction of complex, thin airway trees for pulmonary diagnosis and navigation. The reformulation of segmentation as multi-task connectivity prediction is a targeted architectural response to a known limitation of standard CNNs on tubular structures.

major comments (2)

[Experiments] Experiments (or equivalent results section): no ablation is reported that holds the network architecture, training protocol, and auxiliary input channels (lung distance map + coordinates) fixed while reverting only the output head and loss from 26-connectivity prediction back to standard binary segmentation. Without this control, measured improvements cannot be attributed to the connectivity modeling rather than the added semantic inputs, which directly undermines the central claim that the 26-task formulation enables better perception of airway connectivity.
[Method] Method section (connectivity modeling description): the transformation of binary segmentation into 26 connectivity tasks is presented without a formal definition of how the ground-truth connectivity labels are generated from the binary airway masks or how the final segmentation mask is recovered from the 26 predictions at inference time. This leaves the precise relationship between the auxiliary tasks and the claimed connectivity awareness underspecified.

minor comments (1)

[Abstract] Abstract: the claim of 'superior performance' is stated without naming the datasets, evaluation metrics, or baseline methods, which should be summarized even at this level for a methods paper.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the two major comments point-by-point below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Experiments] Experiments (or equivalent results section): no ablation is reported that holds the network architecture, training protocol, and auxiliary input channels (lung distance map + coordinates) fixed while reverting only the output head and loss from 26-connectivity prediction back to standard binary segmentation. Without this control, measured improvements cannot be attributed to the connectivity modeling rather than the added semantic inputs, which directly undermines the central claim that the 26-task formulation enables better perception of airway connectivity.

Authors: We agree that the requested ablation is required to isolate the contribution of the 26-connectivity formulation. In the revised manuscript we will add this control experiment: the identical network architecture, training protocol, and auxiliary input channels will be retained, but the output head and loss will be changed to standard binary segmentation. Performance differences between this baseline and the full AirwayNet will be reported to attribute gains specifically to connectivity modeling. revision: yes
Referee: [Method] Method section (connectivity modeling description): the transformation of binary segmentation into 26 connectivity tasks is presented without a formal definition of how the ground-truth connectivity labels are generated from the binary airway masks or how the final segmentation mask is recovered from the 26 predictions at inference time. This leaves the precise relationship between the auxiliary tasks and the claimed connectivity awareness underspecified.

Authors: We acknowledge the need for greater formal precision. The revised manuscript will include explicit mathematical definitions: (1) the procedure for generating the 26 ground-truth connectivity labels from each binary airway mask (labeling a voxel's connection to each of its 26 neighbors if both belong to the airway tree), and (2) the inference-time recovery of the binary mask from the 26 predictions (via per-voxel thresholding followed by connectivity-based refinement). These additions will clarify the relationship between the auxiliary tasks and connectivity awareness. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method proposal with no self-referential derivations

full rationale

The paper presents an empirical CNN architecture for airway segmentation that reformulates the task as 26 connectivity predictions plus auxiliary channels. No equations, uniqueness theorems, or fitted parameters are defined in terms of the target outputs. No self-citations are invoked as load-bearing premises. The central claim rests on comparative experimental results rather than any reduction to inputs by construction. This matches the default case of a self-contained method paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that CNNs can learn meaningful connectivity patterns from the reformulated task and that the added distance and coordinate inputs supply useful context; no free parameters or invented entities are specified in the abstract.

axioms (1)

domain assumption CNNs trained on connectivity prediction tasks will capture airway tree structure better than binary segmentation
This is the central modeling choice invoked to justify the 26-task reformulation.

pith-pipeline@v0.9.0 · 5721 in / 1073 out tokens · 20696 ms · 2026-05-24T20:54:11.664782+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

MedI A 36, 52–60 (2017)

Charbonnier, J.P., Van Rikxoort, E.M., Setio, A.A., Scha efer-Prokop, C.M., van Ginneken, B., Ciompi, F.: Improving airway segmentation in computed tomogra- phy using leak detection with convolutional networks. MedI A 36, 52–60 (2017)

work page 2017
[2]

In: MICCAI

C ¸ i¸ cek,¨O., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3D U-Net: learning dense volumetric segmentation from sparse annota tion. In: MICCAI. pp. 424–432. Springer (2016)

work page 2016
[3]

: 3D convolutional neural networks with graph reﬁnement for airway segmentation usin g incomplete data labels

Jin, D., Xu, Z., Harrison, A.P., George, K., Mollura, D.J. : 3D convolutional neural networks with graph reﬁnement for airway segmentation usin g incomplete data labels. In: MLMI. pp. 141–149. Springer (2017)

work page 2017
[4]

In: Image Ana lysis for Moving Or- gan, Breast, and Thoracic Images, pp

Juarez, A.G.U., Tiddens, H., de Bruijne, M.: Automatic ai rway segmentation in chest CT using convolutional neural networks. In: Image Ana lysis for Moving Or- gan, Breast, and Thoracic Images, pp. 238–250. Springer (20 18)

work page
[5]

: Connnet: A long- range relation-aware pixel-connectivity network for sali ent segmentation

Kampﬀmeyer, M., Dong, N., Liang, X., Zhang, Y., Xing, E.P. : Connnet: A long- range relation-aware pixel-connectivity network for sali ent segmentation. IEEE TIP 28(5), 2518–2529 (2019)

work page 2019
[6]

MedIA 14(4), 527–538 (2010)

Lo, P., Sporring, J., Ashraf, H., Pedersen, J.J., de Bruij ne, M.: Vessel-guided airway tree segmentation: A voxel classiﬁcation approach. MedIA 14(4), 527–538 (2010)

work page 2010
[7]

IEEE TMI 31(11), 2093–2107 (2012)

Lo, P., Van Ginneken, B., Reinhardt, J.M., Yavarna, T., De Jong, P.A., Irving, B., Fetita, C., Ortner, M., Pinho, R., Sijbers, J., et al.: Extra ction of airways from CT (exact’09). IEEE TMI 31(11), 2093–2107 (2012)

work page 2093
[8]

In: MICCAI

Meng, Q., Roth, H.R., Kitasaka, T., Oda, M., Ueno, J., Mori , K.: Tracking and segmentation of the airways in chest CT using a fully convolu tional network. In: MICCAI. pp. 198–207. Springer (2017)

work page 2017
[9]

In: MICCAI

Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convoluti onal networks for biomed- ical image segmentation. In: MICCAI. pp. 234–241. Springer (2015)

work page 2015
[10]

In: Proc

Van Rikxoort, E.M., Baggerman, W., van Ginneken, B.: Aut omatic segmentation of the airway tree from thoracic CT scans using a multi-thres hold approach. In: Proc. of Second International Workshop on Pulmonary Image A nalysis. pp. 341– 349 (2009)

work page 2009
[11]

MedIA 24(1), 1–17 (2015)

Xu, Z., Bagci, U., Foster, B., Mansoor, A., Udupa, J.K., M ollura, D.J.: A hybrid method for airway segmentation and automated measurement o f bronchial wall thickness on CT. MedIA 24(1), 1–17 (2015)

work page 2015
[12]

MedIA 51, 13–20 (2019)

Yun, J., Park, J., Yu, D., Yi, J., Lee, M., Park, H.J., Lee, J.G., Seo, J.B., Kim, N.: Improvement of fully automated airway segmentation on v olumetric computed tomographic images using a 2.5 dimensional convolutional n eural net. MedIA 51, 13–20 (2019)

work page 2019

[1] [1]

MedI A 36, 52–60 (2017)

Charbonnier, J.P., Van Rikxoort, E.M., Setio, A.A., Scha efer-Prokop, C.M., van Ginneken, B., Ciompi, F.: Improving airway segmentation in computed tomogra- phy using leak detection with convolutional networks. MedI A 36, 52–60 (2017)

work page 2017

[2] [2]

In: MICCAI

C ¸ i¸ cek,¨O., Abdulkadir, A., Lienkamp, S.S., Brox, T., Ronneberger, O.: 3D U-Net: learning dense volumetric segmentation from sparse annota tion. In: MICCAI. pp. 424–432. Springer (2016)

work page 2016

[3] [3]

: 3D convolutional neural networks with graph reﬁnement for airway segmentation usin g incomplete data labels

Jin, D., Xu, Z., Harrison, A.P., George, K., Mollura, D.J. : 3D convolutional neural networks with graph reﬁnement for airway segmentation usin g incomplete data labels. In: MLMI. pp. 141–149. Springer (2017)

work page 2017

[4] [4]

In: Image Ana lysis for Moving Or- gan, Breast, and Thoracic Images, pp

Juarez, A.G.U., Tiddens, H., de Bruijne, M.: Automatic ai rway segmentation in chest CT using convolutional neural networks. In: Image Ana lysis for Moving Or- gan, Breast, and Thoracic Images, pp. 238–250. Springer (20 18)

work page

[5] [5]

: Connnet: A long- range relation-aware pixel-connectivity network for sali ent segmentation

Kampﬀmeyer, M., Dong, N., Liang, X., Zhang, Y., Xing, E.P. : Connnet: A long- range relation-aware pixel-connectivity network for sali ent segmentation. IEEE TIP 28(5), 2518–2529 (2019)

work page 2019

[6] [6]

MedIA 14(4), 527–538 (2010)

Lo, P., Sporring, J., Ashraf, H., Pedersen, J.J., de Bruij ne, M.: Vessel-guided airway tree segmentation: A voxel classiﬁcation approach. MedIA 14(4), 527–538 (2010)

work page 2010

[7] [7]

IEEE TMI 31(11), 2093–2107 (2012)

Lo, P., Van Ginneken, B., Reinhardt, J.M., Yavarna, T., De Jong, P.A., Irving, B., Fetita, C., Ortner, M., Pinho, R., Sijbers, J., et al.: Extra ction of airways from CT (exact’09). IEEE TMI 31(11), 2093–2107 (2012)

work page 2093

[8] [8]

In: MICCAI

Meng, Q., Roth, H.R., Kitasaka, T., Oda, M., Ueno, J., Mori , K.: Tracking and segmentation of the airways in chest CT using a fully convolu tional network. In: MICCAI. pp. 198–207. Springer (2017)

work page 2017

[9] [9]

In: MICCAI

Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convoluti onal networks for biomed- ical image segmentation. In: MICCAI. pp. 234–241. Springer (2015)

work page 2015

[10] [10]

In: Proc

Van Rikxoort, E.M., Baggerman, W., van Ginneken, B.: Aut omatic segmentation of the airway tree from thoracic CT scans using a multi-thres hold approach. In: Proc. of Second International Workshop on Pulmonary Image A nalysis. pp. 341– 349 (2009)

work page 2009

[11] [11]

MedIA 24(1), 1–17 (2015)

Xu, Z., Bagci, U., Foster, B., Mansoor, A., Udupa, J.K., M ollura, D.J.: A hybrid method for airway segmentation and automated measurement o f bronchial wall thickness on CT. MedIA 24(1), 1–17 (2015)

work page 2015

[12] [12]

MedIA 51, 13–20 (2019)

Yun, J., Park, J., Yu, D., Yi, J., Lee, M., Park, H.J., Lee, J.G., Seo, J.B., Kim, N.: Improvement of fully automated airway segmentation on v olumetric computed tomographic images using a 2.5 dimensional convolutional n eural net. MedIA 51, 13–20 (2019)

work page 2019