Pendekatan Berbasis U-Net untuk Segmentasi Hard Exudate dalam Citra Fundus Retina

Abstract

World Health Organization estimates that globally, 422 million adults over the age of 18 lived with diabetes in 2014. It is also supported by the World Diabetes Foundation estimates that more than 439 million people will be threatened with diabetes by 2030. One of the diseases caused by diabetes is diabetic retinopathy which can cause impaired vision to blindness. Damage to blood vessels and damage to the nerve fibers of the eye are called exudates, which are blood spots containing yellowish-colored fats that have an erratic shape. The types of exudates are divided into two, namely hard exudate and soft exudate. Soft exudate is also known as cotton wool spots and appears with a whitish color with less pronounced edges. While hard exudate occurs due to leakage of proteins and lipid vessels of the retina. The shape appears sharp and bright. To find out where the hard exudate is located in the retinal fundus image, experts or doctors are still looking manually, so it takes a long time to find out location of the hard exudate. Therefore, this research work, contributes to segmenting hard exudate using deep learning, which the method is U-Net. The final result of hard exudates segmentation using U-Net methods is validated with ground truth images by measuring accuracy, sensitivity, and specificity metric score. The results of hard exudate segmentation show for the U-Net metric score is 0.993, 0.454, and 0.997 respectively.