A Comparison of Transfer Learning Techniques, Deep Convolutional Neural Network and Multilayer Neural Network Methods for the Diagnosis of Glaucomatous Optic Neuropathy

Early glaucoma diagnosis prevents permanent structural optic nerve damage and consequent irreversible vision impairment. Longitudinal studies have described both baseline structural and functional factors that predict the development of glaucomatous change in ocular hypertensive and glaucoma suspects. Although there is neither a gold standard for disease diagnosis nor progression, photographic assessment of the optic nerve head remains a mainstay in the diagnosis and management of glaucoma suspects and glaucoma patients. We describe a method aimed at both detecting pathologic changes, characteristic of glaucomatous optic neuropathy in optic disc images, and classification of images into categories glaucomatous/suspect or normal optic discs. Three different deep-learning algorithms used are transfer learning, deep convolutional neural network, and deep multilayer neural network that extract features automatically based on clinically relevant optic-disc features. Of the total of 455 cases extracted from the RIM-ONE public dataset (version 2), consisting of 348 training, 87 validation and 20 test cases, the proposed approach classified images with a training accuracy of 98.16%. We hypothesise that this approach can support the clinical decision algorithm in the diagnosis of glaucomatous optic neuropathy.