Optic Disk Detection Using Feature Clustering and Classification in Retinal Fundus Images

An initial but critical task in automated clinical decision support for retinal pathology diagnosis is optic disc localization. Optic disk detection and localization in fundus imaging requires robust solutions that can mitigate potential large intra-class variations due to camera blurring, eye motion, and the presence of pathologies. We have combined several feature calculations with a supervised classification scheme to detect optic disk locations in a series of retinal images that contain normal and diseased optical features. In our experimentation, we perform blood vessel segmentation with parabola fitting to create accumulator matrices that find maximal intersections, which discriminate optical disk areas from regions of similar intensities. We also employ a novel set of edge histograms to further reduce false detections in areas where blood vessel configurations cause multiple areas of high vessel intersections. We then reduce the problem space by calculating wavelet decomposition and hierarchical clustering prior to supervised classification stages. This step significantly decreases computation time and increases detection accuracy. The detection methodology is evaluated using the publically available STARE retinal fundus dataset. The results display high discriminative classification capabilities using the feature sets on multiple trained classifiers, leading to highly specific optic disk detection accuracy.