Eye Microvessels

Conjunctival vessel segmentation and measurement

A U-Net segments the microvessels on a slit-lamp photo of the eye; each vessel is then skeletonised, classified as large or medium and measured in micrometres (F-score 0.78).

Try it

Choose one of the presets or upload an image

Results will appear here
Uploading…

Algorithm

Stack
  • Python 3.10
  • FastAPI
  • MLFlow
  • Pytorch
  • segmentation_models.pytorch
  • Albumentations
  • OpenCV
  • Shapely
Dataset
  • 1088 slit-lamp images 1624x1232
  • 570 with GeoJSON vessel markup
  • Pre-training on CHASE_DB1 (28 fundus images)
  • Structural markup method
Training
  • UNet + ResNet50
  • 2 classes (background / vessel)
  • F-score 0.78
  • Skeleton-based width and length in µm

Model

Pipeline scheme