Pyramid-Net

Pyramid-Net: Intra-Layer Pyramid-Scale Retinal Vessel Segmentation Network

Pyramid-Net performs retinal vessel segmentation using intra-layer pyramid-scale feature aggregation to enhance delineation of thin vessel structures in fundus images.

Key Features:

• Intra-Layer Pyramid-Scale Feature Aggregation: Utilizes Intra-Layer Pyramid-Scale Aggregation Blocks (IPABs) to generate higher- and lower-scale branches within each layer for multi-scale feature fusion.
• Pyramid Inputs Enhancement: Optimizes input data processing to improve segmentation accuracy.
• Deep Pyramid Supervision: Applies supervision at multiple network levels to improve multi-scale learning and generalization.
• Pyramid Skip Connections: Enables efficient inter-layer information flow and feature integration with reduced computational cost.

Scientific Applications:

• Retinal Vessel Analysis: Segments thin retinal vessels in fundus photography to support early detection of diabetic retinopathy and glaucoma; validated on DRIVE, STARE, and CHASE-DB1 datasets.

Methodology:

Pyramid-Net aggregates multi-scale features within each network layer via IPABs that create scale-specific branches, combined with pyramid inputs enhancement, deep pyramid supervision, and pyramid skip connections to improve segmentation accuracy and computational efficiency.