Intelligent feature selection technique for segmentation of fundus images

We propose a technique for selecting effective features for fundus image segmentation. The technique has made it possible to conduct smart feature analysis with the aid of color subspaces when solving a problem of selecting the regions of interest. The relevance of the problem is associated with enhancing the effectiveness of laser coagulation surgery. The technique is based on the texture analysis of image patterns. Source texture features are derived from different statistical image descriptors computed using the MaZda library (image histogram, image gradient, run-length matrix, and the co-occurrence matrix). The analysis of the information content of the resulting feature space and selection of most effective features is conducted using the discriminant data analysis. We determine the best size of an image fragmentation block for segmentation the ocular fundus and feature sets enabling the regions of interest to be identified with a required accuracy when analyzing the following four classes of images: exudates, thick vessels, thin vessels, and intact areas. The feature selection technique utilizes K-means-based clustering whereas the Euclidean and Mahalanobis distance is used as a similarity measure. The required minimal size of the fragmentation block and the similarity measure are chosen based on the criterion of the minimal clustering error among all smallest block sizes. The proposed technique enables not only the informative features to be extracted in particular color spaces but also the most informative color subspace to be identified.