Peer Reviewed

1

DOI

10.17077/omia.1010

Conference Location

Boston, MA, USA

Publication Date

September 2014

Abstract

The earliest sign of the diabetic retinopathy is the appearance of small red dots in retinal fundus images, designated by microaneurysms. In this paper a scale-space based method is proposed for the microaneurysms detection. Initially, the method performs a segmentation of the retinal vasculature and defines a global set of microaneurysms candidates, using both coarser and finer scales. Using the finer scales, a set of microaneurysms candidates are analysed in terms of shape and size. Then, a set of gaussian-shaped matched filters are used to reduce the number of false microaneurysms candidates. Each candidate is labeled as a true microaneurysm using a new neighborhood analysis method. The proposed algorithm was tested with the training Retinopathy Online Challenge (ROC) dataset, revealing a 47% Sensitivity with an average number of 37.9 false positives per image.

Rights

Copyright © 2014, Ivo Soares, Miguel Castelo-Branco and Antonio M.G. Pinheiro.

Included in

Ophthalmology Commons

Share

COinS
 
Sep 14th, 12:00 AM Sep 14th, 12:00 AM

Microaneurysms detection using a novel neighborhood analysis

Boston, MA, USA

The earliest sign of the diabetic retinopathy is the appearance of small red dots in retinal fundus images, designated by microaneurysms. In this paper a scale-space based method is proposed for the microaneurysms detection. Initially, the method performs a segmentation of the retinal vasculature and defines a global set of microaneurysms candidates, using both coarser and finer scales. Using the finer scales, a set of microaneurysms candidates are analysed in terms of shape and size. Then, a set of gaussian-shaped matched filters are used to reduce the number of false microaneurysms candidates. Each candidate is labeled as a true microaneurysm using a new neighborhood analysis method. The proposed algorithm was tested with the training Retinopathy Online Challenge (ROC) dataset, revealing a 47% Sensitivity with an average number of 37.9 false positives per image.