VascuConNet: an enhanced connectivity network for vascular segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Medical image segmentation is crucial for disease localization and quantification.
• Enhancing segmentation continuity is a key challenge in clinical applications.
• Retinal vessel segmentation aids in diagnosing various eye conditions.

Purpose of the Study:

• To develop a novel segmentation model for improved retinal vessel segmentation accuracy.
• To leverage vessel orientation, boundary, and continuity constraints for enhanced segmentation.
• To address the challenge of segmentation continuity in medical imaging.

Main Methods:

• A cascaded U-Net and long-short-term memory network (LSTM) architecture was employed.
• An orientation information enhancement module using orientation convolution was introduced.
• A hybrid loss function combining connectivity, boundary, and cross-entropy losses was designed.

Main Results:

• The proposed model demonstrated excellent segmentation performance.
• High accuracy was achieved across three standard retinal vessel segmentation datasets (CHASE_DB1, DRIVE, ARIA).
• The integration of orientation and continuity constraints significantly improved segmentation outcomes.

Conclusions:

• The novel segmentation model effectively enhances retinal vessel segmentation continuity and accuracy.
• The developed method shows significant potential for clinical applications in ophthalmology.
• The hybrid loss function and orientation enhancement module are key contributors to the model's success.