3D augmented fundus images for identifying glaucoma via transferred convolutional neural networks

Glaucoma is a chronic and irreversible retinopathy threatening the vision of millions of patients around the world. Its early diagnosis and treatment can help to prolong the period of sight deterioration from no visual impairment to blindness, whereas the screening and diagnosis of glaucoma in clinical remains challenging because some key assessment criteria like cup-to-disc ratio is limited by subjective analysis and intra- and inter-observer variability. This paper exploits the potential of new augmented image data of the optic nerve head (ONH) combining with the latest deep learning networks to achieve better diagnosis of glaucoma.

This paper explores the potential value of additional three-dimensional topographic map of the optic nerve head proceeded by the latest deep learning approaches, i.e. convolutional neural networks to improve the diagnosis efficiency. Specifically, 3D topography map of the ONH and RGB fundus image has been used to train the transferred AlexNet and VGG-16 networks. The diagnostic performance is compared to those achieved by using the 2D fundus images only.

The 3D topographic map of ONH reconstructed from the shape from shading method provides better visualization of the structure of optic cup and disc. These new enhanced dataset was employed to train the proposed deep learning networks and finally achieve diagnostic accuracy of 94.3% which is superior to the networks trained via 2D conventional images.

Employing the deep learning neural networks with augmented 3D images can increase the accuracy of automatic separating glaucoma and non-glaucoma fundus images. It may be used as an objective tool in developing computer assisted diagnosis systems for assessment of glaucoma.