GKE-TUNet: Geometry-Knowledge Embedded TransUNet Model for Retinal Vessel Segmentation Considering Anatomical...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Automated retinal vessel segmentation is vital for diagnosing eye conditions like retinopathy.
• Deep learning models struggle with intricate vascular structures and small vessels in dense retinal regions.

Purpose of the Study:

• To introduce a novel segmentation model, GKE-TUNet, that embeds retinal vessel anatomical topology.
• To improve the accuracy of automated retinal vessel segmentation, especially in challenging vascular patterns.

Main Methods:

• A skeleton extraction network pre-trains to capture vessel topology.
• A graph attention network (GAT) processes topological features within a TransUNet architecture.
• Graph features are fused with dense feature maps to enhance segmentation.

Main Results:

• The GKE-TUNet model demonstrated competitive performance on DRIVE, CHASE-DB1, and STARE datasets.
• The method effectively extracts complex intertwined structures and subtle small vessels.
• Integration of topological knowledge improved segmentation accuracy.

Conclusions:

• The GKE-TUNet model offers a robust approach for automated retinal vessel segmentation.
• Embedding explicit anatomical knowledge enhances deep learning model performance for medical imaging tasks.
• This method holds promise for advancing computer-aided diagnosis and retinopathy screening.