AirMorph: Topology-Preserving Deep Learning for Pulmonary Airway Analysis

TL;DR

AirMorph is introduced, a robust, end-to-end deep learning pipeline enabling fully automatic and comprehensive airway anatomical labeling at lobar, segmental, and subsegmental resolutions that can be used to create digital atlases of the lung.

Abstract

Accurate anatomical labeling and analysis of the pulmonary structure and its surrounding anatomy from thoracic CT is getting increasingly important for understanding the etilogy of abnormalities or supporting targetted therapy and early interventions. Whilst lung and airway cell atlases have been attempted, there is a lack of fine-grained morphological atlases that are clinically deployable. In this work, we introduce AirMorph, a robust, end-to-end deep learning pipeline enabling fully automatic and comprehensive airway anatomical labeling at lobar, segmental, and subsegmental resolutions that can be used to create digital atlases of the lung. Evaluated across large-scale multi-center datasets comprising diverse pulmonary conditions, the AirMorph consistently outperformed existing segmentation and labeling methods in terms of accuracy, topological consistency, and completeness. To simplify clinical interpretation, we further introduce a compact anatomical signature quantifying critical morphological airway features, including stenosis, ectasia, tortuosity, divergence, length, and complexity. When applied to various pulmonary diseases such as pulmonary fibrosis, emphysema, atelectasis, consolidation, and reticular opacities, it demonstrates strong discriminative power, revealing disease-specific morphological patterns with high interpretability and explainability. Additionally, AirMorph supports efficient automated branching pattern analysis, potentially enhancing bronchoscopic navigation planning and procedural safety, offering a valuable clinical tool for improved diagnosis, targeted treatment, and personalized patient care.

Figures (27)

• Figure 1: Overview of the AirMorph's development, fine-grained evaluation, and clinical applications. a) Model development. AirMorph is a fully automated framework for extracting subsegmental anatomical bronchi from thoracic CT scans. It comprises three stages: (1) binary airway modeling from CT scans, (2) feature extraction from a graph-based representation of the airway tree, and (3) anatomical labeling based on branch-wise features. b) Evaluation. AirMorph supports unified and fine-grained evaluation of both binary airway modeling and anatomical labeling via graph node-level performance metrics. c)-d) Clinical applications. AirMorph facilitates the real-world clinical applications. c) It enables efficient analysis of airway branching patterns across the entire bronchial tree. d) Fine-grained airway signatures quantify structural abnormalities between patient cohorts with lung disease and healthy controls.
• Figure 2: Detailed information of the datasets for AirMorph. a) A total of 3,023 thoracic computed tomography (CT) scans were collected from four distinct institutions and partitioned into a primary dataset and independent test datasets. ATM++ serves as the primary dataset, comprising 620 patients with detailed annotations that include both binary airway labels and subsegmental-level anatomical labels. An additional 806 patients from the LIDC-IDRI dataset and 1,597 patients from the NLST dataset were also included in this study. b) The fine-grained annotation includes the binary airway from CT scans, along with five lobar anatomies, nineteen kinds of segmental anatomies, and one hundred and twenty-seven subsegmental anatomies. Additionally, the first row represents the airway anatomies of patients with mild conditions, wherea the last row depicts structural alterations observed in patients with advanced pulmonary fibrosis.
• Figure 3: Fine-grained evaluation of AirMorph. a-b) Detailed evaluation of the tree length detection rate of AirMorph compared with UNet and nnUNet at the subsegmental branch level. a) compares selected bronchi in the left upper lobe: $LB^{{1+2}^{a}}$, $LB^{{1+2}^{b}}$, $LB^{{1+2}^{c}}$, and $LB^{{3}^{a}}$. b) shows corresponding comparisons for the right lower lobe: $RB^{{8}^{a}}$, $RB^{{8}^{b}}$, $RB^{{9}^{a}}$, and $RB^{{9}^{b}}$. c) Graphnode-wise anatomical labeling performance across representative bronchial regions. Comparisons are shown for the primary dataset (c1-c2) and the external independent test set (c3-c4), involving AirMorph, GCN, and TNN. d) Confusion matrices illustrating subsegmental labeling accuracy on the primary dataset across different models. e-f) Ablation Study on external independent test sets. e) Total number of reconstructed airway branches. f) Cumulative tree length of the predicted airway structures. g) Subsegmental labeling consistency for adjacent bronchi. h) Topological distance of predicted airway graphs. P-values are specified as $\ast$p < 0.05, $\ast\ast$p < 0.01, $\ast\ast\ast$p < 0.001, $\ast\ast\ast\ast$p < 0.0001, n.s, not significant.
• Figure 4: Comprehensive characterization of airway branching patterns across lobar, segmental, and subsegmental levels using AirMorph. a) Distribution of branching pattern types across all five lobes, from lobar to subsegmental level, showing the prevalence of one-stem, two-stem, bifurcated, trifurcated, and co-trunk configurations. b) Take the $LB^{1+2}$ as a representative segmental branch, and visualize the airway branching pattern statistics along with their anatomical mapping through 3D endoscopic views. c) Inter-segmental branching pattern analysis among $RB^{1}$, $RB^{2}$, and $RB^{3}$, visualized with annotated 3D bronchoscopic views. d) Inter-segmental pattern analysis of $LB^{4}$ and $LB^{4}$, showing the prevalence of independent and co-trunk configurations.
• Figure 5: AirwaySignature representation and morphological comparison across disease types. a) Definition of six airway morphological descriptors computed from anatomically labeled branches using AirMorph: stenosis, ectasia, tortuosity, divergence, length, and complexity. b) Visualization of the AirwaySignature matrix in a representative case with atelectasis. b1 shows the corresponding CT scan highlighting the lesion regions happen in the right middle and lower lobe. b2 displays anatomical branch-wise airway signatures distributions. Color encoding indicates statistical deviation from the healthy reference distribution. c) Disease-level comparison of AirwaySignature profiles across five pulmonary conditions (pulmonary fibrosis, emphysema, atelectasis, consolidation, and reticular opacities) and healthy controls. Distinct morphological trends are visible among subgroups, supporting the discriminative power of AirwaySignature.