An evaluation of performance measures for arterial brain vessel segmentation

Area of Science:

• Medical image analysis
• Cerebrovascular imaging
• Computational anatomy

Background:

• Cerebral vessel segmentation is crucial for extracting clinical information from the vascular tree.
• Current methods lack a standardized performance measure for evaluating segmentation quality.
• This hinders the development and validation of new segmentation techniques.

Purpose of the Study:

• To develop a framework for selecting the most suitable performance measure for arterial brain vessel segmentation.
• To identify a standardized metric that accurately reflects segmentation quality.

Main Methods:

• Simulated non-overlapping segmentation errors in magnetic resonance angiography data from 10 patients.
• Manual visual scoring of approximately 300 segmentation variations per patient.
• Correlation analysis (Spearman) between visual scores and rankings from 22 common performance measures.

Main Results:

• Distance-based measures, specifically balanced average Hausdorff distance (rank 1) and average Hausdorff distance (rank 2), showed the highest correlation with manual rankings.
• Overlap-based measures like the Dice coefficient ranked lower (rank 7).

Conclusions:

• Average Hausdorff distance-based measures are recommended as the standard for evaluating cerebral vessel segmentation.
• These measures effectively identify segmentation errors, particularly in high-quality segmentations.
• Adoption of these measures can accelerate the development of advanced vessel segmentation approaches.