Automated Diagnosis of Ocular Fundus Images

Abstract

 

Purpose

 

To develop a novel method of automated fundus image diagnosis algorithm using an error-correcting output code (ECOC), an ensemble learning technique, and compare its performance to other learning/pattern recognition algorithms.

 

Methods

 

We collected an image dataset which covered 11 retinal diagnoses plus normal. A computer-aided design (CAD) system using ECOC was developed to assign fundus images automatically into 12 classes (Central retinal vein occlusion, branch retinal vein occlusion, nonproliferative diabetic retinopathy, proliferative diabetic retinopathy, choroidal neovascularization, subretinal neovascularization, geographic atrophy, drusen, retinitis, presumed ocular histoplasmosis syndrome, Coats’ disease, and normal). The images were taken with a Topcon 50EX camera with 2400x1400 pixel resolution. To represent different aspects of abnormalities and diagnoses, 120 color fundus images (10 samples per class), were selected by two retina specialists. The images were labeled to evaluate the CAD system that was developed with learning algorithms. In the classifier architecture, the process levels were hidden, extracting statistical features the classifier found relevant for the classification task. Training and testing were done with 10-fold cross-validation. Figure 1 shows steps of CAD.

 

Results

 

Our computer algorithm which used an error-correcting output code with Linear Discrimant Analyses (LDA) features successfully identified the 11 diagnoses, with recognition rates from 25 to 42. Figure 2 compares the recognition rate of different baseline methods with the ECOC classifier.

 

Conclusions

 

Automated image recognition of fundus images can provide a rapid tool in helping in the diagnosis of retinal pathology. This is the first attempt for automatic retinal diagnosis out of multiple classes. Large numbers of feature selection/extraction and classifiers have been applied to the problem in order to obtain a reasonable performance. As classification performance is improved and as the number of retinal pathologies included in the algorithm is increased, the significance of applying image classification has the potential to help clinicians in the diagnosis and management of their patients.

 

 

 

(ECOC: error-correcting output code, LDA: linear discriminant analysis, MLD: multilayer perceptron, PCA: principal component analysis, SVM: support vector machine)

 

Keywords: 550 imaging/image analysis: clinical • 549 image processing • 688 retina