CHASE-AV retinal vessel reference dataset

A public retinal vessel reference dataset CHASE-AV made available by Kingston University, London in collaboration with St. George’s, University of London. CHASE-AV is built upon the publicly available CHASE_DB1 dataset. This is a subset of retinal images of multi-ethnic children from the Child Heart and Health Study in England (CHASE) dataset. This subset contains 28 retinal images captured from both eyes from 14 of the children recruited in the study. In this subset each retinal image is also accompanied by one ground truth image. This is provided in the form of manual arteriole/venule segmentation maps made by an expert observer under clinical guidance, in which image pixels are colour coded with red = arterioles, blue = venule, green = crossings, white = uncertain, and black = background. Making this subset publicly available allows for the scientific community to train and test computer vision algorithms (specifically arteriole/venule segmentation methodologies). Most importantly this subset allows for performance comparisons - several algorithms being evaluated on the same database allows for direct comparisons of their performances to be made.

Image

The Child Heart and Health Study in England (CHASE) surveyed cardiovascular health in 200 primary schools across London, Birmingham, and Leicester. Retinal imaging in 46 schools showed links between vessel tortuosity and early cardiovascular risk in over 1000 ethnically diverse British children. Images from both eyes were taken with a handheld Nidek NM-200-D fundus camera at 30° FOV and 1280×960 resolution, later cropped to 999×960. The images typically have uneven illumination, low vessel contrast, and arterioles with a bright central reflex. A subset of these images formed the CHASE_DB1 retinal vessel reference dataset, now extended to CHASE-AV. It includes 28 images from both eyes of 14 children (8 white, 3 South Asian, 3 other ethnicities; mean age 10) from one North-East London school. Each image has an expert-annotated arteriole/venule segmentation ground truth.