A Hierarchical Scheme for Geodesic Anatomical Labeling of Airway Trees

Aasa Feragen, Jens Petersen, Megan Owen, Pechin Lo, Laura H. Thomsen, Mathilde M. W. Wille, Asger Dirksen, Marleen de Bruijne

Introduction

Quantitative analysis of lung disease progression using Computed Tomography (CT) requires the precise identification and comparison of specific airway branches across longitudinal scans or different patient cohorts. However, manual anatomical labeling is a labor-intensive task prone to significant inter-observer variability. This challenge is compounded by "anatomical noise"—the inherent variability in branching patterns between individuals—and structural distortions caused by pathologies like Chronic Obstructive Pulmonary Disease (COPD).

The Challenge of Anatomical Variability

Automated labeling methods must contend with topological noise, such as spurious branches resulting from segmentation errors or missing branches due to poor image quality. As shown in Figure 2, labeling involves identifying specific segments (e.g., R1, R2, R3) within larger structures like the Right Upper Lobe (RUL). Previous automated methods often struggled when these structures were distorted or when branching patterns deviated from the standard anatomical atlas.

Anatomical labeling of the right upper lobe (RUL) and its segments (R1, R2, R3) on a 3D airway model.

Geodesic Labeling in Tree-Space

To address these issues, the researchers employed a "tree-space" framework. In this mathematical model, an entire airway tree is represented as a single point in a high-dimensional, non-Euclidean space. The distance between two points—the geodesic—quantifies the cost of morphing one tree's geometry and topology into another. This allows the algorithm to treat anatomical variations as smooth transitions, making it naturally resistant to noise.

The proposed algorithm utilizes two primary strategies:

  1. Hierarchical Search: To manage the exponential complexity of matching a 20-label tree, the process is decomposed into a top-down hierarchy. It first identifies the trachea and main bronchi before proceeding to lobar and segmental levels.
  2. Centerline Robustness: By focusing exclusively on the tree's skeleton (centerlines) rather than airway thickness or volume, the method remains robust against the morphological changes typical of severe lung disease.
Conceptual representation of tree-space morphing and hierarchical labeling of airway branches.

Results and Clinical Significance

The algorithm was validated on a dataset of 40 airway trees from 20 subjects, including healthy individuals and patients with varying degrees of COPD. Key findings include:

  • High Accuracy: The method achieved 88.35% agreement with expert consensus in cases where experts agreed with each other.
  • Superior Reproducibility: The algorithm's consistency (82.59%) exceeded the inter-expert agreement (79.48%), suggesting it provides a more reliable baseline for longitudinal studies.
  • Pathology Independence: Accuracy did not degrade with increased COPD severity, confirming the utility of centerline-based features for clinical populations.
  • Efficiency: With a processing time of approximately 5 minutes per tree on a standard laptop, the tool is practical for clinical integration.

Limitations and Future Work

The current hierarchical approach is "greedy"; an error at a higher level (e.g., the main bronchus) propagates to all descendant branches. Furthermore, the algorithm currently attempts to assign all 20 standard labels even if a branch is anatomically absent due to surgery or disease. Future iterations may incorporate probabilistic labeling to allow the algorithm to identify missing or anomalous structures rather than forcing a match.

Summary updated

Abstract

We present a fast and robust supervised algorithm for labeling anatomical airway trees, based on geodesic distances in a geometric tree-space. Possible branch label configurations for a given tree are evaluated based on distances to a training set of labeled trees. In tree-space, the tree topology and geometry change continuously, giving a natural way to automatically handle anatomical differences and noise. The algorithm is made efficient using a hierarchical approach, in which labels are assigned from the top down. We only use features of the airway centerline tree, which are relatively unaffected by pathology. A thorough leave-one-patient-out evaluation of the algorithm is made on 40 segmented airway trees from 20 subjects labeled by 2 medical experts. We evaluate accuracy, reproducibility and robustness in patients with Chronic Obstructive Pulmonary Disease (COPD). Performance is statistically similar to the inter- and intra-expert agreement, and we found no significant correlation between COPD stage and labeling accuracy.
BibTeX
@inproceedings{feragen2012hierarchical,
  author = {Aasa Feragen and Jens Petersen and Megan Owen and Pechin Lo and Laura H. Thomsen and Mathilde M. W. Wille and Asger Dirksen and Marleen de Bruijne},
  title = {{A Hierarchical Scheme for Geodesic Anatomical Labeling of Airway Trees}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2012},
  doi = {10.1007/978-3-642-33454-2\_19},
}
Figure