Deep learning-driven pulmonary artery and vein segmentation reveals demography-associated vasculature anatomical...

Area of Science:

• Medical Imaging
• Radiology
• Computational Anatomy

Background:

• Pulmonary artery-vein segmentation is vital for diagnosing lung diseases and planning surgeries.
• Current methods often use Computed Tomography Pulmonary Angiography (CTPA), which involves contrast agents and associated health risks.
• Accurate segmentation using non-contrast CT has been a significant challenge due to lower image quality and vessel conspicuity.

Purpose of the Study:

• To develop a novel method for accurate pulmonary artery-vein segmentation in both non-contrast CT and CTPA.
• To enable non-invasive and safer medical imaging analysis for pulmonary vasculature.
• To facilitate large-scale population studies on pulmonary vessel anatomy and its clinical correlations.

Main Methods:

• Proposed High-abundant Pulmonary Artery-vein Segmentation (HiPaS), integrating spatial normalization and iterative segmentation.
• Leveraged lower-level vessel segmentations as priors for higher-level segmentations.
• Trained and validated on a multi-center dataset of 1073 CT volumes with manual annotations.

Main Results:

• Achieved superior segmentation performance with a Dice score of 91.8% and sensitivity of 98.0%.
• Demonstrated non-inferiority of HiPaS on non-contrast CT compared to CTPA.
• Enabled analysis of 11,784 participants, revealing associations between vessel abundance and demographic/clinical factors.

Conclusions:

• HiPaS provides an accurate and robust method for pulmonary artery-vein segmentation using non-contrast CT.
• This approach offers a safer, non-invasive alternative for clinical diagnostics and anatomical research.
• The method supports large-scale studies, advancing understanding of pulmonary vascular health.