Multi-Granularity Topological Reasoning for Anatomically Consistent Vasculature Parsing.

Area of Science:

• Ophthalmology and Medical Imaging

Background:

• Accurate retinal vascular segmentation is crucial for diagnosing systemic diseases and guiding clinical decisions.
• Challenges in segmenting retinal arteries and veins (A/V) include variations in vessel size, low contrast, and anatomical variability, particularly in ultra-wide field (UWF) imaging.
• Existing methods often produce fragmented and topologically inconsistent segmentation results.

Purpose of the Study:

• To develop a novel deep learning network, Ultra, for precise A/V segmentation in fundus images.
• To improve the topological accuracy and consistency of automated retinal vascular segmentation.
• To address the limitations of current methods in handling UWF imaging complexities.

Main Methods:

• Proposed Ultra, a cascaded two-stage network comprising PriorNet for multi-scale vascular priors and RefineNet for topology-aware refinement.
• Introduced the neighboring pixel connectivity regularization (NICER) layer with a connectivity prediction union (CPU) module for enhanced topological coherence.
• Utilized a pixel-wise local connectivity loss as auxiliary supervision to reinforce structural reasoning.

Main Results:

• Ultra achieved state-of-the-art performance in A/V segmentation and topological preservation on UWF fundus imaging datasets.
• Demonstrated excellent generalization capabilities on conventional color fundus photography (CFP) datasets.
• The proposed NICER layer and connectivity loss effectively improved topological consistency.

Conclusions:

• Ultra significantly enhances the precision and topological integrity of retinal A/V segmentation, especially in challenging UWF images.
• The network's robustness and broad applicability are confirmed by its strong performance on diverse fundus imaging datasets.
• Ultra represents a significant advancement for quantitative retinal vascular analysis in both research and clinical settings.