Coronary artery segmentation framework based on three types of U-Net and voting ensembles

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate coronary artery (CA) segmentation is vital for diagnosing cardiovascular diseases.
• Existing segmentation methods struggle with CA complexity, leading to discontinuities and false positives.

Purpose of the Study:

• To propose a robust voting-based ensemble segmentation framework for improved CA segmentation.
• To enhance the capture of global and local CA structural features.

Main Methods:

• Utilized a voting-based ensemble of three U-Net variants: a lightweight U-Net for global context, and patch-based and multi-slice U-Nets for local details.
• Integrated segmentation results from the three models using a voting strategy to achieve the final segmentation map.

Main Results:

• The proposed framework achieved a Dice score of 82.31% on a large dataset.
• Demonstrated improved segmentation by effectively handling CA structural complexity and reducing pseudo-CA regions.

Conclusions:

• The developed ensemble segmentation framework offers a significant advancement in coronary artery segmentation accuracy.
• This approach holds promise for more reliable computer-aided diagnosis of coronary artery diseases.