Peer Reviewed

1

DOI

10.17077/omia.1032

Conference Location

Munich, Germany

Publication Date

October 2015

Abstract

With the growing diabetes epidemic, retina specialists have to examine a tremendous amount of fundus images for the detection and grading of diabetic retinopathy. In this study, we propose a first automatic grading system for diabetic retinopathy. First, a red lesion detection is performed to generate a lesion probability map. The latter is then represented by 35 features combining location, size and probability information, which are finally used for classification. A leave-one-out cross-validation using a random forest is conducted on a public database of 1200 images, to classify the images into 4 grades. The proposed system achieved a classification accuracy of 74.1% and a weighted kappa value of 0.731 indicating a significant agreement with the reference. These preliminary results prove that automatic DR grading is feasible, with a performance comparable to that of human experts.

Rights

Copyright © 2015 Lama Seoud, Jihed Chelbi and Farida Cheriet

Included in

Ophthalmology Commons

Share

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

Automatic Grading of Diabetic Retinopathy on a Public Database

Munich, Germany

With the growing diabetes epidemic, retina specialists have to examine a tremendous amount of fundus images for the detection and grading of diabetic retinopathy. In this study, we propose a first automatic grading system for diabetic retinopathy. First, a red lesion detection is performed to generate a lesion probability map. The latter is then represented by 35 features combining location, size and probability information, which are finally used for classification. A leave-one-out cross-validation using a random forest is conducted on a public database of 1200 images, to classify the images into 4 grades. The proposed system achieved a classification accuracy of 74.1% and a weighted kappa value of 0.731 indicating a significant agreement with the reference. These preliminary results prove that automatic DR grading is feasible, with a performance comparable to that of human experts.