ABSTRACT
162Since diabetic retinopathy (DR) is an ocular disease with a fast progression to blindness, prevention and early diagnosis are of utmost importance. Fast action in the early stages of the disease can be decisive in preserving the patient's vision. This disease is characterized by the appearance of a heterogeneous set of lesions, among which microaneurysms and hemorrhages are some of the most important ones. Detection and classification of these lesions is a part of DR classification. In this chapter, we present the approach and results of our experiments with automated detection of such lesions with a density clustering approach, by integrating preprocessing, superpixelation, and density clustering, and by targeting classification of individual lesions. The approach is based on preprocessing, simple linear iterative clustering (SLIC) and density-based spatial clustering (DBSCAN), filtering, and classification of the retina background images. After detailing the approach, we analyze it experimentally, in particular, we study its achievements in terms of accuracy, sensitivity, and specificity when different classifiers are employed. Finally, we reach important conclusions and delineate future work on the issue.
