Coronary Artery Vascular Segmentation on Limited Data via Pseudo-Precise Label

Area of Science:

• Medical imaging analysis
• Computer vision
• Artificial intelligence in healthcare

Background:

• Accurate segmentation of medical coronary artery angiograms is crucial but challenging.
• High manual labeling costs and network accuracy trade-offs limit traditional semantic segmentation methods.
• The unique characteristics of cardiac angiography data necessitate novel approaches.

Purpose of the Study:

• To develop a cost-effective method for improving coronary artery segmentation accuracy.
• To introduce a novel approach for generating 'pseudo-precise' labels.
• To enhance the performance of segmentation networks while minimizing labor costs.

Main Methods:

• Proposed a new method for generating 'pseudo-precise' labels.
• Developed a complementary training pipeline utilizing these pseudo-precise labels.
• Evaluated the method's impact on segmentation performance using f1-score metrics.

Main Results:

• The proposed method significantly reduces the cost associated with manual data labeling.
• Achieved an increase in the f1-score ranging from 4% to 11% compared to baseline methods.
• Demonstrated improved network performance with reduced labor investment.

Conclusions:

• The 'pseudo-precise' labeling and training pipeline offer a viable solution for medical image segmentation.
• This approach effectively balances network accuracy and the reduction of manual annotation efforts.
• It presents a significant advancement for automated analysis of coronary artery angiograms.