Connectivity-based Segmentation of Retinal Vessels in Eye Fundus Images

A new unsupervised method for segmentation of objects of diverse nature with the common feature of connectivity (e.g. branching trees or net-shaped objects) is proposed. A preferred application to the vasculature segmentation of retinal images has been illustrated using images from DRIVE database. In the pre-processing stage, the method overcomes the common problem of non-uniform illumination of eye fundus images. The method follows with an iterative algorithm that starts with a seed and adds, at each step, a new vessel segment connected to the previously segmented part. The result preserves the connectivity as a distinct feature of the retinal vessel tree. The segmentation performance is evaluated through common signal detection metrics: sensitivity, specificity and accuracy.