Vessel Analysis of Retinal Images

The retinal vasculature is a primary biomarker for vision-threatening and systemic disease. Our research develops methods to extract, classify, and assess retinal vessels from fundus and angiography images — combining topology-aware deep learning, multi-modal integration, interactive refinement, and generative modeling to produce vessel maps that are accurate, anatomically consistent, and clinically useful.

Research directions

  • Structure-aware vessel segmentation. Beyond pixel-level classification, our models explicitly learn vessel connectivity and multi-scale structure, producing segmentations that preserve the topological integrity of the vascular tree.
  • Artery/vein classification. Topology-aware and multi-modal approaches that propagate A/V labels along connected vessel segments, fusing fundus and FA evidence for more reliable classification.
  • Multimodal and FA-assisted vessel extraction. Leveraging registration with fluorescein angiography to supervise fundus-based vessel segmentation at a finer scale than standalone fundus data permits.
  • Interactive and hybrid assessment. Integrating deep segmentation with interactive editing, graphical modeling, and multimodal prompt-based refinement for precise clinical vessel assessment.
  • Generative modeling for vessel data. Using diffusion models to synthesize structurally realistic fundus images with paired vessel annotations, addressing data scarcity for rare conditions.
  • Image enhancement for downstream vessel analysis. Deep learning frameworks that improve fundus image quality — correcting illumination, blur, and artifacts — to make vessel analysis more reliable.