Ensemble Based Optic Disc Detection Method

Abstract

Diabetic retinopathy is one of leading causes of preventable blindness in the world. Detection of the optic disc from fundus photographs is an important part in developing automatic diabetic retinopathy screening systems. In this paper we present a new method for optic disc detection. The method combines different optic disc detection algorithms into an ensemble. The method starts by applying each optic disc detection algorithm to the input image, which produces an optic disc location probability map for each of the methods applied. The outputs of each algorithm are combined by weighting their respective probability maps. The weights are found using the simulated annealing search algorithm. In the new probability map, the optic disc location is chosen by taking the point in the probability map with the highest probability level. The new method was tested using publicly available DRIVE, DiaretDB0, DiaretDB1 and DRiDB datasets. The method outperformed other state-of-the art algorithms used in creation of the ensemble and the results obtained show that the optic disc can be robustly detected in images from both healthy patients and patients, which have a lot of visible artifacts.