A Neural Network Based Optic Disc Segmentation

Identifying the pixels belonging to the optic disc region accurately and segmenting the optic disc's ambiguous boundary in retinal images is crucial in developing systems that assist in detecting retinal diseases. Such segmentation is an arduous task to accomplish manually and eased off by computer-aided systems. The proposed method makes effective use of a convolutional neural network that utilizes the benefits of dropout regularization, skip connections, and a minimal number of fully connected layers. The network captures the context and localization information accurately. It is an image input-image output network model that is tolerant of a wide range of input image resolutions and field of views available in the datasets. The method is tested on four standard retinal image datasets, DIARETDB1, DIARETDB0, MESSIDOR, and DRIONS-DB, and achieved an accuracy of 99.26%, 99.53%, 97.86%, and 92.50%, respectively. A comparative analysis of the proposed algorithm has segmented the optic disc better than many existing algorithms.