Automated generation of directed graphs from vascular segmentations

Area of Science:

• Medical Imaging Informatics
• Computational Anatomy
• Graph Theory Applications

Background:

• Automated feature extraction from medical images is crucial for quantitative analysis.
• Existing segmentation tools provide semi-automated vascular segmentations.
• Accurate identification of vascular substructures is challenging.

Purpose of the Study:

• To develop and validate a graph-based technique for automated vascular substructure identification.
• To extract morphological features and perform accurate diameter measurements from vascular trees.
• To improve the efficiency and accuracy of pulmonary vascular analysis.

Main Methods:

• Utilized a 3D parallel thinning algorithm to generate vascular skeletons.
• Transformed skeletons into directed graphs with bifurcations and endpoints as nodes.
• Employed machine-learning classifiers to prune false vascular structures.
• Computed least-squares cubic splines for centerline path analysis.
• Performed semantic labeling of graph components with pulmonary anatomy.

Main Results:

• Achieved high accuracy (⩾0.97) in semantic labeling of pulmonary anatomy.
• Demonstrated high correlation (r⩾0.77) between automated and manual diameter measurements.
• Successfully extracted morphological features using graph-based centerline paths.
• Effectively pruned false vascular structures using machine learning.

Conclusions:

• The proposed graph-based technique enables accurate automated feature extraction from vascular segmentations.
• This method facilitates precise morphological analysis and diameter quantification of pulmonary arteries.
• The approach holds significant potential for advancing quantitative medical imaging analysis.