Expert-Guided Knowledge Distillation for Semi-Supervised Vessel Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Blood vessel segmentation is crucial for medical diagnosis and surgery.
• Existing methods often require extensive manual annotations, hindering development.
• Complex vascular structures present significant segmentation challenges.

Purpose of the Study:

• To propose a novel semi-supervised vessel segmentation framework, EXP-Net, to address the limitations of data annotation.
• To enhance knowledge distillation by incorporating an expert network within a semi-supervised learning approach.
• To improve the accuracy of blood vessel segmentation in medical images, particularly in challenging cases.

Main Methods:

• Developed EXP-Net, a semi-supervised vessel segmentation framework based on the Mean Teacher model.
• Introduced an expert network with knowledge and connectivity enhancement modules for improved feature modeling.
• Utilized a vision transformer for long-range dependency modeling and non-parametric skeletonization for topological and geometric enhancement.

Main Results:

• EXP-Net demonstrates state-of-the-art performance in blood vessel segmentation.
• The framework effectively handles weak vessel connectivity and poor pixel contrast.
• Achieved superior results on subcutaneous vessel, retinal vessel, and coronary artery datasets.

Conclusions:

• EXP-Net offers a practical solution for vessel segmentation with limited labeled data.
• The proposed expert network enhances knowledge distillation for more robust segmentation.
• This framework advances the field of medical image analysis for clinical applications.