• Osareh, A.
• Mirmehdi, Majid
• Markham, R.
• Thomas, B.

Abstract

Aim: to develop an automatic method to analyse colour retinal imagesto detect and classify exudates (EXs) in diabetic maculopathy.Method: We propose a method comprising four different stages.Firstly, colour retinal images are normalised due to wide variationin the colour of retina from different patients. Then, to improveboth the contrasting attributes of EXs and the overall coloursaturation in the image, a local contrast enhancement technique isapplied. In the third stage, a colour segmentation method based onFuzzy C-means (FCM) clustering is performed. In our experiments, FCMcould distinguish more than 98\% of EXs successfully. In the finalstage, ten relevant features are employed to divide the feature spaceinto two disjoint classes and then the FCM segmented regions areclassified as EX/non-EX classes by a Neural Network (NN) classifierusing the feature measurements. Thenetwork training and testing wasperformed on 42 retinal images, which contained 4037 objects, each ofwhich was labeled by an ophthalmologist.Results: The NN could achieve 92\% sensitivity and 82\% specificity.However, alternative results can be obtained by varying the thresholdon the network output, e.g. 83\% sensitivity and 94\% specificity. Inaddition, we investigated other classifiers and compare the results.The overall performances for the NN, K-Nearest Neighbors, RadialBasis Function and Quadratic Gaussian classifiers were 90.1\%,86.32\% (K=4), 87.39\% and 78.33\% respectively.Conclusion: The results are very promising and show that automatedidentification of EX lesions on the basis ofcolour information isof practical use to ophthalmologists. We are presently obtaining moredata and expect the performance to be improved continually throughricher training and testing information.

Topics

• impedance spectroscopy
• experiment
• clustering