CGNet

CGNet segments retinal vasculature in optical coherence tomography angiography (OCTA) images to generate vessel segmentation maps for diagnosis and analysis of retinal diseases.

Key Features:

• End-to-End Three-Stage Architecture: An end-to-end three-stage architecture structures coarse, fine, and refining stages for progressive vessel segmentation.
• Channel and Position Attention (CPA) modules: Employs channel and position attention (CPA) modules to enhance vessel feature extraction from OCTA images.
• Graph Reasoning Network (GRN): Uses a graph reasoning network (GRN) to model structural relationships in retinal vasculature.
• U-shaped encoder-decoder: Implements a U-shaped neural network encoder-decoder in the coarse stage to integrate CPA and GRN for feature extraction.
• Vessel confidence maps: Generates vessel confidence maps in the coarse stage to highlight potential vascular structures.
• 3-channel composite merging: Forms a 3-channel composite by merging the confidence map, the original OCTA image, and a fine image map to refine micro-vasculatures.
• Refining stage fusion: Fuses refined and fine images in the final refining stage to produce the ultimate segmentation result.
• Validated performance: Reported validation on public datasets with area under the ROC curve (AUC) values of 94.29% and 85.62% across two datasets.

Scientific Applications:

• Retinal OCTA vessel segmentation: Performs segmentation of retinal vasculature in OCTA images for quantitative and qualitative analysis.
• Diagnosis and monitoring of retinal diseases: Supports identification of pathological vascular changes to assist diagnosis, early detection, and treatment planning for retinal diseases.

Methodology:

CGNet executes three explicit computational stages: a coarse stage that combines CPA and GRN within a U-shaped encoder-decoder to produce vessel confidence maps; a fine stage that merges confidence maps with the original OCTA image and a fine image map into a 3-channel composite to refine micro-vasculature; and a refining stage that fuses refined and fine images to generate the final segmentation.