Publication

Abstract : Relating to diagnosis of ophthalmologic diseases, retinal fundus images provide valuable clinical information. Retinal blood vessel analysis gives diagnostic information about contraction of the retinal nerve fiber layer and variation in the appearance of the optic nerve head, which leads to the development of reproducible glaucomatous visual field defects. Here, a novel retinal vessel detection method with disc region detection is proposed from fundus images using multilabel residual convolutional neural network architecture with skip connection and up sampling. In the training phase, data augmentation was performed to improve training performance. We use joint loss function with adaptive gradient descent optimizer to locate the blood vessels in the disc region. Efficacy of the proposed method has been experimented on DRIVE and STARE databases to analyze vessels in the vicinity of disc region for better clinical diagnosis of retinal diseases and its progression. The proposed method attains an accuracy of 84.90% for DRIVE database and 79.80% for STARE database.