Fundus Images Segmentation by Unsupervised Classification

In this paper, we present an original unsupervised segmentation scheme which splits a grey level image into different sets of connected pixels whose grey levels are homogeneous. This approach is based on an analysis of a triangular table denoted ”Normalized connectivity degrees pyramid”. This method is used in order to detect cytomegalovirus retinitis lesions by fundus image analysis. First, we determine the number of pixels classes and their cores. The core of each class Cj is represented by an interval of grey levels [minCj ;maxCj ]. For classification purpose, the pixels whose grey level belongs to such an interval are labelled to the corresponding class. The other pixels are assigned by comparison of their conditional probability to belong to the different classes.