Automatic Segmentation of Exudates in Ocular Images using Ensembles of Aperture Filters and Logistic Regression

Marco Benalcázar,Marcel Brun,Virginia Ballarin

doi:10.1088/1742-6596/477/1/012021

Abstract

Hard and soft exudates are the main signs of diabetic macular edema (DME). The segmentation of both kinds of exudates generates valuable information not only for the diagnosis of DME, but also for treatment, which helps to avoid vision loss and blindness. In this paper, we propose a new algorithm for the automatic segmentation of exudates in ocular fundus images. The proposed algorithm is based on ensembles of aperture filters that detect exudate candidates and remove major blood vessels from the processed images. Then, logistic regression is used to classify each candidate as either exudate or non-exudate based on a vector of 31 features that characterize each potensial lesion. Finally, we tested the performance of the proposed algorithm using the images in the public HEI-MED database.

Highlights

Exudates are one the main signs of diabetic macular edema (DME), which occurs when the retina swells as a complication of diabetic retinopathy
The ultimate goal of the automatic segmentation of exudates is the diagnosis of the DME condition
The area under the receiver operating characteristic (ROC) curve (AUC) is 0.77. This value and the shape of the ROC curve indicate that our algorithm tends to detect exudates in some images that belong to healthy patients

Summary

Introduction

Exudates are one the main signs of diabetic macular edema (DME), which occurs when the retina swells as a complication of diabetic retinopathy. A mass screening of all diabetic patients, even of those experiencing no vision issues, would help to diagnose DME early enough for optimal treatment [1]. At this time, such an undertaking would be too work and time intensive because each image must be analyzed by a specialist, making the early diagnosis of this pathology difficult. We combine tools from digital image processing, pattern recognition, and machine learning to propose a new algorithm for the automatic segmentation of both hard and soft exudates.

Images and Methodology

Aperture Filters

Results and Discussion