Skeleton2Mask: Skeleton-supervised airway segmentation

Abstract

Airway segmentation has achieved considerable success. However, it still hinges on precise voxel-wise annotations, which are not only labor-intensive and time-consuming but also subject to challenges like missing branches, discontinuous branch labeling, and erroneous edge delineation. To tackle this, this paper introduces two novel contributions: a skeleton annotation (SKA) strategy for airway tree structures, and a sparse supervision learning approach — Skeleton2Mask, built upon SKA for dense airway prediction. The SKA strategy replaces traditional slice-by-slice, voxel-wise labeling with a branch-by-branch, control-point-based skeleton delineation. This approach not only enhances the preservation of topological integrity but also reduces annotation time by approximately 80%. Its effectiveness and reliability have been validated through clinical experiments, demonstrating its potential to streamline airway segmentation tasks. Nevertheless, the absolute sparsity of this annotation, along with the typical tree structure, can easily cause the failure of sparse supervision learning. To tackle this, we further propose Skeleton2Mask, a two-stage label propagation learning method, involving dual-stream buffer propagation and hierarchical geometry-aware learning, to ensure reliable and structure-friendly dense prediction. Experiments reveal that 1) Skeleton2Mask outperforms other sparsely supervised approaches on two public datasets by a large margin, achieving comparable results to full supervision with no more than 3% of airway annotations. 2) With the same annotation cost, our algorithm demonstrated significantly superior performance in both topological and voxel-wise metrics.