Automated segmentation of hard exudates based on matched filtering

Abstract

In 2015, according to the International Diabetes Federation (IDF), around 415 million of people worldwide lived with diabetes and it was predicted to be increased by 642 million of people in 2040. One of the diabetes complications that affect the retina is known as diabetic retinopathy (DR). It is indicated by the presence of hard exudates as the main pathology of DR. In retinal fundus images, hard exudates appear as bright lesion which has some similar characteristics with the optic disc. This paper proposes a method to automatically detect hard exudates. At first, the green channel is extracted from the retinal colour fundus image. The complement of green channel is used to increase the contrast between hard exudates and the background. The complemented image is filtered by using matched filter. Optic disc (OD) is detected based on initial optic disc enlargement in L band of HSL colour space. Afterwards, optic disc is removed from filtered image to obtain the candidates of hard exudates followed by the morphological operation. The proposed method is validated by using 60 colour fundus images from DIARETDB1 dataset. The final results of segmented exudates are verified by comparing with their ground truth images. The average level of accuracy, sensitivity and specificity achieved are 99.99%, 90.38% and 99.99%, respectively. These results indicate that the proposed method successfully detected the hard exudates. Hence, it is recommended to be implemented as a part of DR grading system development.