Semantic Segmentation in Human Retinal Images

Abstract

Three leading causes of blindness in the world are diabetic retinopathy, glaucoma and macular degeneration. Vision loss due to these diseases can easily be avoided by early diagnosis and treatment. Many systemic and ocular diseases manifest themselves in retina, which is transparent in appearance making the in-vivo examination possible. An automated mass screening system based on retinal images can reduce the risk of blindness and vision loss. In this regard, the fundamental step is accurate segmentation of retinal pathologies (i.e., exudates, hemorrhages, and cotton-wool spots) and retinal landmarks (i.e. retinal blood vessels and optic disc). In this thesis an end-to-end framework based on encoder decoder fully convolutional neural network is proposed for simultaneous segmentation of retinal pathologies and landmarks. The proposed network is inspired from U-Net, with introduction of Batch Normalization layers and a down-sampling block to helps in segmentation of small retinal lesions. For the purpose of evaluation, a new public retinal image dataset is created from Messidor dataset, where the pixel-level multi-class annotations are done by the trained observers and verified by ophthalmologist. The proposed technique is evaluated on DRIVE, HRF, eOphtha and Messidor datasets, and it achieved remarkable performance in retinal anatomical structure segmentation using standard evaluation metrics.