Please use this identifier to cite or link to this item:
Title: Hard exudate detection in retinal fundus images using supervised learning
Authors: Nipon Theera-Umpon
Ittided Poonkasem
Sansanee Auephanwiriyakul
Direk Patikulsila
Keywords: Computer Science
Issue Date: 1-Jan-2019
Abstract: © 2019, Springer-Verlag London Ltd., part of Springer Nature. The patients with diabetes have a chance to develop diabetic retinopathy (DR) which affects to the eyes. DR can cause blindness if the patients do not control diabetes. The patients with DR will have an impairment of metabolism of glucose causing a high glucose level in blood vessel called hyperglycemia. It leads to abnormal blood vessel and ultimately results in leakage of blood or fluid like lipoproteins, which are deposited under macular edema called hard exudates. They are normally white or yellowish-white with margins. Hard exudates are often arranged in clumps or circinate rings and located in the outer layer of the retina. The aim of this research was to detect hard exudates by applying several image processing techniques and classify them by using supervised learning methods including support vector machines and some neural network approaches, i.e., multilayer perceptron (MLP) network, hierarchical adaptive neurofuzzy inference system (hierarchical ANFIS), and convolutional neuron networks. DIARETDB1 which contains 89 fundus images is exploited as a dataset for evaluation. Hard exudate candidates are extracted by morphological techniques and classified by the classifiers trained by extracted patches with the corresponding ground truths. The tenfold cross-validation is applied to assure the generalization of the results. The proposed method achieves the area under the curve (AUC) of 0.998 when the MLP network is applied. The AUCs for all four classifiers are more than 0.95. This shows that the combination of image processing techniques and suitable classifiers can perform very well in hard exudate detection problem.
ISSN: 14333058
Appears in Collections:CMUL: Journal Articles

Files in This Item:
There are no files associated with this item.

Items in CMUIR are protected by copyright, with all rights reserved, unless otherwise indicated.