On feature extraction techniques for eye images

Abstract

Optical Coherence Tomography (OCT) is a non-invasive method which can cap­ture high-definition images of cross-section of the retina. Based on the thickness of different layers of the retina in OCT images, one can diagnose ocular diseases in an early stage. Numerous algorithms, including deep learning methods, have been used by various researchers on OCT images for retinal layer segmentation of normal eyes and eyes with ocular diseases. Performances of deep learning algorithms have been compared with the state-of-the-art algorithm/software at the time of research. In this thesis, retinal segmentation of OCT images is carried out for nine boundaries, equiva­lent to segmenting eight retinal layers. A thresholding optimization method using gra­dient filter is first proposed for the segmentation problem. Different neural network architecture combining U-net and other CNNs have also been investigated. To reduce network complexity, a method is proposed based on the concept of domain decom­position when training a large volume of data on cloud platform. At last, 3D CNN classification model and 3D U-net segmentation model are proposed and trained to classify an eye as healthy or glaucomatous and segment the eight retinal layers using 3D volumetric OCT eye scans.