Accurate Airway Tree Segmentation in CT Scans via Anatomy-Aware Multi-Class Segmentation and Topology-Guided...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Intrathoracic airway segmentation in computed tomography (CT) is crucial for analyzing respiratory diseases like COPD, asthma, and lung cancer.
• Current deep learning methods struggle with complete airway tree segmentation due to low contrast, noise, complex topology, and incomplete labeling, limiting diagnostic accuracy.

Purpose of the Study:

• To develop an anatomy-aware, multi-class airway segmentation method enhanced by topology-guided iterative self-learning to address challenges in airway segmentation.
• To improve the completeness and accuracy of airway segmentation, particularly for deeper, peripheral branches.

Main Methods:

• An anatomy-aware multi-class segmentation task is formulated to handle intra-class imbalance in airway data.
• An iterative self-learning scheme with a novel breakage attention map and topology-guided pseudo-label refinement is proposed to address incomplete labeling.
• The method iteratively refines pseudo-labels by connecting broken airway branches to achieve higher sensitivity.

Main Results:

• The proposed method achieved top performance in the EXACT'09 and ATM'22 challenges.
• Significant improvements were observed on public and private datasets, with at least 6.1% more detected tree length and 5.2% more tree branches extracted.
• Comparable precision was maintained while enhancing segmentation completeness.

Conclusions:

• The developed anatomy-aware, topology-guided iterative self-learning method effectively addresses challenges in intrathoracic airway segmentation.
• This approach enhances the completeness and accuracy of airway segmentation, offering significant benefits for respiratory disease analysis and computer-aided diagnosis.