SURVS: A Swin-Unet and game theory-based unsupervised segmentation method for retinal vessel

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Manual annotation of medical images, particularly vascular structures, is labor-intensive and expensive.
• Developing unsupervised methods for vessel segmentation is crucial for efficient disease detection and analysis.

Purpose of the Study:

• To design an unsupervised retinal vessel segmentation model using Swin-Unet and game theory.
• To overcome the limitations of manual annotation in medical image analysis.
• To improve the accuracy and efficiency of retinal vessel segmentation for disease detection.

Main Methods:

• Constructed extreme pseudo-mapping functions and derived pseudo-masks using binary segmentation and mathematical morphology.
• Developed a Swin-Unet based model to find the optimal mapping function, incorporating an Image Colorization proxy task.
• Proposed a game filter inspired by game theory (Neighbor's Collision game) to refine segmentation masks and reduce errors.

Main Results:

• The unsupervised model demonstrated effectiveness on DRIVE, STARE, and CHASE_DB1 datasets.
• The proposed method showed significant advantages over traditional image segmentation and other unsupervised models.
• The integration of Swin-Unet and game theory yielded accurate retinal vessel segmentation.

Conclusions:

• The developed unsupervised model offers a valuable alternative to manual annotation for retinal vessel segmentation.
• The combination of Swin-Unet, pseudo-mapping, and game theory provides a robust framework for medical image analysis.
• This approach has the potential to enhance early disease detection through improved medical image segmentation.