Image-based classification of diabetic retinopathy using machine learning

In this paper we present experimental results of an automated method for image-based classification of diabetic retinopathy. The method is divided into three stages: image processing, feature extraction and image classification. In the first stage we have used two image processing techniques in order to enhance their features. Then, the second stage reduces the dimensionality of the images and finds features using the statistical method of principal component analysis. Finally, in the third stage the images are classified using machine learning algorithms, particularly, the naive Bayes classifier, neural networks, k-nearest neighbors and support vector machines. In our experimental study we classify two types of retinopathy: non-proliferative and proliferative. Preliminary results show that k-nearest neighbors obtained the best result with 68.7% using f-measure as metric, for a data set of 151 images with different resolutions.