Peer Reviewed

1

DOI

10.17077/omia.1035

Conference Location

Munich, Germany

Publication Date

October 2015

Abstract

In many diseases with a cardiovascular component, the geometry of microvascular blood vessels changes. These changes are specific to arteries and veins, and can be studied in the microvasculature of the retina using retinal photography. To facilitate large-scale studies of artery/vein-specific changes in the retinal vasculature, automated classification of the vessels is required. Here we present a novel method for artery/vein classification based on local and contextual feature analysis of retinal vessels. For each vessel, local information in the form of a transverse intensity profile is extracted. Crossings and bifurcations of vessels provide contextual information. The local and contextual features are integrated into a non-submodular energy function, which is optimized exactly using graph cuts. The method was validated on a ground truth data set of 150 retinal fundus images, achieving an accuracy of 88.0% for all vessels and 94.0% for the six arteries and six veins with highest caliber in the image.

Rights

Copyright © 2015 Koen Eppenhof, Erik Bekkers, Tos T.J.M. Berendschot, Josien P.W. Pluim, and Bart M. ter Haar Romeny

Included in

Ophthalmology Commons

Share

COinS
 
Oct 9th, 12:00 AM Oct 9th, 12:00 AM

Retinal Artery/Vein Classification via Graph Cut Optimization

Munich, Germany

In many diseases with a cardiovascular component, the geometry of microvascular blood vessels changes. These changes are specific to arteries and veins, and can be studied in the microvasculature of the retina using retinal photography. To facilitate large-scale studies of artery/vein-specific changes in the retinal vasculature, automated classification of the vessels is required. Here we present a novel method for artery/vein classification based on local and contextual feature analysis of retinal vessels. For each vessel, local information in the form of a transverse intensity profile is extracted. Crossings and bifurcations of vessels provide contextual information. The local and contextual features are integrated into a non-submodular energy function, which is optimized exactly using graph cuts. The method was validated on a ground truth data set of 150 retinal fundus images, achieving an accuracy of 88.0% for all vessels and 94.0% for the six arteries and six veins with highest caliber in the image.