Structure and Position-Aware Graph Neural Network for Airway Labeling

Weiyi Xie · January 1, 2022

Paper Published   Medical Imaging Analysis   Available Online
Qries

Table of Contents

Motivation
Method
Results
Cite

Motivation

CT imaging is an excellent tool for in vivo quantitative airway analysis. Such an analysis can be performed efficiently when an automatic airway segmentation is available and is often applied regionally for specific anatomical branches. An automated airway labeling can expedite such a process. Airway labeling is also useful to plan bronchoscopy interventions, where the location of bronchoscopic can be automatically tagged by the labeled airway branches.

Method

The highlight of this method is that I use positional encodings of airway branches in the network training and inference such that the learned node representation is not only structure-aware (like other methods based on graph neural networks) but also position-aware. The overview of our method is shown in the following figure.

Qries

please refer to the published paper for more details.

Results

  • We evaluated this algorithm from the COPDGene study (n=220) using 5-fold cross-validation. The overall branch classification accuracy reaches 91.18\%, where the baseline approach is at 83.83\%.

  • t-SNE on how learned branch features look like in 2D and how they are distributed across different anatomical labels in the CNN method and the proposed method including the learned positional encodings.

    Qries

Cite

If you find this algorithm useful in your research, please consider citing:

@article{Xie22,
    author={Xie, Weiyi and Jacobs, Colin and Charbonnier, Jean-Paul and van Ginneken, Bram},
    title={Structure and position-aware graph neural network for airway labeling},
    journal={arXiv preprint arXiv:2201.04532},
    year={2022} 
}

Twitter, Facebook