Automatic Detection of Eye Retinal Pathologies

This thesis aims to design and implement a system that automatically detects retinal eye pathologies. The retina is an essential part of the eye that, as the only organ in the body, contains light-sensitive cells that make a vision possible. For the treatment of eye disease to be successful, early detection and precise examination of its extent are crucial. The proposed system based on the supplied image automatically generates masks representing occurrences of individual pathologies. The result is then presented to the user. A convolutional neural network based on the U-Net architecture handles the evaluation. The network was trained on the Indian Diabetic Retinopathy Image Dataset (IDRiD), which contains 81 images of the retina and associated annotations. The system was evaluated using the AUC-PR score (area under the precision-recall curve). Segmentation of hard exudates, soft exudates, hemorrhages and microaneurysms achieved an AUC-PR score of 74%, 50%, 45% and 33%, respectively. This work proposes an innovative architecture that, if further developed, has the potential to be used by ophthalmologists for diagnosing and determining the extent of retinal disease.

bioinformatics, human eye, eye retina, image processing, segmentation, eye retinal pathologies, machine learning, neural networks