Bronchovascular Tree-Guided Weakly Supervised Learning Method for Pulmonary Segment Segmentation.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Machine Learning in Medicine

Background:

• Pulmonary segment segmentation is vital for cancer diagnosis and surgical planning.
• Manual segmentation is time-consuming and difficult due to indistinct segment boundaries.

Purpose of the Study:

• To develop a novel weakly supervised learning method for accurate pulmonary segment segmentation.
• To reduce the need for laborious voxel-wise annotations by incorporating anatomical knowledge.

Main Methods:

• Anatomy-Hierarchy Supervised Learning (AHSL) uses hierarchical supervision at lobe and segment levels.
• A dual-level supervision paradigm constrains segments within lobes and uses sparse bronchovascular annotations.
• A consistency loss encourages smooth boundaries, and a two-stage strategy utilizes bronchovascular information for segmentation.

Main Results:

• AHSL achieved high Dice coefficients (0.923-0.934) for artery and bronchus segmentation on CT scans.
• External dataset validation showed excellent generalization and clinical validity with high radiologist scores (3.71/3.69).
• The method outperforms state-of-the-art techniques without requiring voxel-wise annotations.

Conclusions:

• AHSL offers superior pulmonary segment segmentation by integrating hierarchical supervision, consistency loss, and a two-stage strategy.
• The method demonstrates clinical validity and excellent generalization capabilities.
• AHSL significantly advances automated lung segmentation for clinical applications.