Pulmonary lobe segmentation based on ridge surface sampling and shape model fitting.

Area of Science:

• Medical Imaging
• Radiology
• Computational Anatomy

Background:

• Quantitative analysis of lung lobes in computed tomography (CT) is crucial for characterizing diseases like chronic obstructive pulmonary disease (COPD).
• Accurate lung lobe segmentation is often hindered by the unavailability of airway and vessel segmentations, which typically aid in defining lobe boundaries.

Purpose of the Study:

• To develop a fully automatic algorithm for lung lobe segmentation in CT scans, specifically designed to work without relying on airway or vessel structures.
• To provide a method for quantitative analysis of lung lobes, particularly beneficial for complex lung diseases.

Main Methods:

• Utilized ridge surface image features and a particle system to identify potential fissure locations based on Hessian matrix analysis.
• Employed principal component analysis (PCA) for modeling shape variations of lobe boundaries and fitting these models to candidate fissure locations.
• Applied thin plate spline (TPS) interpolation to create final lung lobe boundaries using identified particle points.

Main Results:

• The algorithm demonstrated performance comparable to pulmonologist-generated segmentations on 100 CT scans (50 inspiratory, 50 expiratory) from the COPDGene study.
• Achieved high accuracy, with only 5.85% of lobes showing Dice scores below 0.9, and showed good results even with accessory fissures, incomplete fissures, emphysema, and low-dose protocols.
• Boundary identification accuracy was highest near detectable fissures on CT scans.

Conclusions:

• The developed algorithm effectively performs lung lobe segmentation without auxiliary structures like vessels and airways.
• The method shows good performance even in challenging cases with incomplete or absent fissures and high image noise, though these remain the most difficult scenarios.