Blood vessels segmentation in nonmydriatic images using wavelets and statistical classifiers

This work describes a new framework for automatic analysis of optic fundus nonmydriatic images, focusing on the segmentation of the blood vessels by using pixel classification based on pattern recognition techniques. Each pixel is represented by a feature vector composed of color information and measurements at different scales taken from the continuous wavelet (Morlet) transform as well as from mean and order filtering applied to the green channel. The major benefit resulting from the wavelet application to the optic fundus images is its multiscale analysing capability in tuning to specific frequencies, thus allowing noise filtering and blood vessel enhancement in a single step. Supervised classifiers are then applied to label each pixel as either a vessel or a nonvessel. Two different strategies to select the training set have been devised: (1) the blood vessels of a sample image are completely drawn by hand, leading to a labeled image (i.e. vessels × nonvessel pixels) which is used to train the classifier, to be applied to other images; (2) the vessels located in a given small portion of the target image are drawn by hand and the remaining fundus image is segmented by a classifier trained using the hand-drawn portion to define the training set. The latter strategy is particularly suitable for the implementation of a semiautomated software to be used by health workers in order to avoid the need of setting imaging parameters such as thresholds. Both strategies have been extensively assessed and several successful experimental results using real-case images have been obtained.