Purpose:
To automatically segment the blood vessels on color fundus imageand measure the vessel width.
Methods:
This method is a graph-based approach. The two boundaries ofthe same blood vessel are segmented simultaneously by convertingthe two-boundary segmentation problem into a two-slice, three-dimensionalsurface segmentation problem by wrapping the vessel segmentfrom the centerline. The vessel centerline is generated froma vessel probability image, which is a prior work of our group.The two-slice, three-dimensional surface segmentation problemis further converted into the problem of computing a minimumclosed set problem in a node-weighted graph. The cost functionis generated from the orientation sensitive first order derivativeof Gaussian of the green channel. This method is tested on theREVIEW database. The analysis of the relationship between vesselwidth and the distance to the optic disc is also done on a setof 600 fundus images.
Results:
One result image is given Figure 1. The analysis of the relationshipbetween vessel width and the distance to the optic disc is givenin Figure 2. The vessel width shows a monotonic decrease fromthe optic disc to image edge. The black circle shows a distanceof 400 pixels to the optic disc.
Conclusions:
This is a novel method to measure retinal vessel width accurately.The algorithm shows a performance that is comparable to humanobservers. This method shows a good performance especially onfine vessels, which are usually hard to measure because of lowimage contrast and noise.
Keywords: image processing • imaging/image analysis: non-clinical