MSO-GP: 3-D segmentation of large and complex conjoined tree structures.

Area of Science:

• Computer Vision
• Computational Biology
• Medical Imaging

Background:

• Segmenting large, complex 3D tree structures is challenging, especially in computational biology with limited data for learning algorithms.
• Existing methods struggle with accuracy and efficiency for tasks like separating arteries and veins in pulmonary CT angiograms.

Purpose of the Study:

• To develop an unsupervised method for robust segmentation of complex 3D tree structures.
• To address the challenge of artery-vein separation in non-contrast pulmonary CT angiography.
• To provide a foundational step for vascular geometry measurement and pulmonary disease diagnosis.

Main Methods:

• Developed a novel unsupervised visual geometry approach combining multiscale opening with geodesic path propagation (MSO-GP).
• Utilized skeletonization of the conjoined structure to guide geodesic path propagation for improved segmentation.
• Applied digital topology and morphometry principles, circumventing the need for supervised learning.

Main Results:

• The MSO-GP method demonstrated robust segmentation of complex tree structures.
• Successfully achieved artery-vein separation from non-contrast pulmonary CT angiograms.
• Outperformed competing methods in terms of segmentation time and reduced user intervention on synthetic, pig lung, and human lung data.

Conclusions:

• The MSO-GP method offers an effective, unsupervised solution for challenging 3D tree structure segmentation.
• This technique is crucial for accurate vascular geometry measurement, aiding in pulmonary disease diagnosis and the development of image-based phenotypes.
• The approach shows significant promise for improving efficiency and accuracy in medical image analysis workflows.