Vitreous Hemorrhage Segmentation in Fundus Images by Using an Efficient-UNet Network

Eye exams based on fundus image are used to detect abnormalities in the retina and optic nerve. These exams are able to detect issues such as diabetic retinopathy, retinal detachment, glaucoma, ocular melanoma, thrombosis, hemorrhages, among others. Fundus image examinations are also used to detect diseases such as macular degeneration, diabetes, hypertension, and coagulation disorders associated with vitreous hemorrhage, which is a common issue in individuals over the age of 50. Usually, Computer vision and artificial intelligence techniques have been used to support some diagnosis process associated to vitreous hemorrhage. However, there are still challenges during the fundus image analysis and lesions detection, these are due to the position, shape, and image color obtained from fundus examination. In this work, an automatic vitreous hemorrhages segmentation strategy is presented, which improve the well know segmentation techniques based on image thresholding. For doing that, fundus images are binarized by applied an segmentation image thresholding technique, which produces a subset of images. These binarized images are used as inputs to train an fully convolutional neural network. Our CCN architecture is based on Efficient-UNet network. The performance was evaluated by using a set of 200 images extracted from the IDRID dataset, an open access set of fundus images examination. The obtained results reports a rate of precision of 85.93%, compared to the 56% achieved by using a classic segmentation image thresholding strategy.