Skeleton2Mask: Skeleton-supervised airway segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Accurate airway segmentation is crucial but hindered by laborious voxel-wise annotations.
• Existing methods struggle with missing branches, discontinuous labeling, and inaccurate boundaries.

Purpose of the Study:

• To develop an efficient annotation strategy and a robust learning method for dense airway prediction using sparse data.
• To reduce the time and effort required for airway annotation while improving segmentation accuracy.

Main Methods:

• Introduced a novel skeleton annotation (SKA) strategy using branch-by-branch control points, reducing annotation time by ~80%.
• Developed Skeleton2Mask, a two-stage sparse supervision learning method with dual-stream buffer propagation and hierarchical geometry-aware learning.

Main Results:

• Skeleton2Mask significantly outperforms other sparse methods, achieving results comparable to full supervision with only 3% of annotations.
• The proposed method demonstrates superior topological and voxel-wise performance compared to existing algorithms at the same annotation cost.

Conclusions:

• The SKA strategy and Skeleton2Mask model offer a highly efficient and accurate solution for airway segmentation.
• This approach has the potential to streamline clinical workflows by drastically reducing annotation burden.