Structure-aware retinal image segmentation for ophthalmic and brain-health applications

Retinal image segmentation is a key step towards quantitative ophthalmic analysis, supporting vessel morphology measurement, lesion assessment, and biomarker discovery. However, reliable segmentation remains difficult because the blood vessels are thin and topology-sensitive, while fundus lesions are small, irregular and imbalanced. Dense pixel-level annotations are also expensive, which makes weakly supervised segmentation an attractive alternative approach.

In this talk, I will present the research on efficient and structure-aware retinal image segmentation. I will discuss problems with noise in the weakly supervised pipelines and I will introduce structured pseudo-supervision refinement as a strategy for improving pseudo-mask quality before segmentation training. I will then show the results of experiments on two large datasets (OCTA-500 and DDR) and discuss the strengths and limitations of my approach. Finally, I will outline future work on structure-aware refinement, image enhancement, and retinal biomarker extraction for ophthalmic and brain-health applications.