Abstract
Optical coherence tomography (OCT) is a non-invasive and non-contact imaging modality with micrometer resolution; therefore, it is the standard-of-care imaging modality in ophthalmology. OCT is used to image the anterior segment including the cornea, and the posterior segment of the eye, including the retina. Several eye diseases are diagnosed using segmentation-based thickness maps, which capture the changes caused by different diseases to the anatomical structures of the corneal and retinal layers. In this dissertation, we propose image processing and deep learning methods for the segmentation and classification of corneal and retinal OCT images. The segmentation methods are used for the detection of the corneal and retinal layer interfaces which are important for the generation of thickness maps. The classification methods are used to directly diagnose corneal and retinal diseases using OCT images without the need of performing segmentation. In the first part of the dissertation, we propose model-based and graph-based segmentation methods for the detection of corneal layer interfaces in OCT images. The proposed model-based method use random sample consensus (RANSAC) method with second-order polynomial models to detect the corneal layer interfaces. This method can reliably segment six corneal layer interfaces with regular shapes. However, it requires the OCT images to have sufficiently high signal-to-noise ratio (SNR) to succeed. Therefore, we propose a multi-stage graph-based method that can detect the variability in the corneal layer interfaces where it does not use a prior model for the interfaces. The graph-based method use a directed graph with gradient, directional, and multiplier costs. The graph-based method can segment five corneal layer interfaces. In the second part of the dissertation, we propose multi-resolution classification networks to diagnose corneal and retinal diseases using OCT images. We compare the performance of the proposed networks with existing networks using receiver operating characteristic (ROC) curves. We visualize the networks to show the learned features and gain insight how they work. Also, we propose multi-resolution segmentation networks for the detection of corneal and retinal layer interfaces. We compare the proposed networks with existing segmentation networks. We visualize the output segmentation of all networks to visually compare their performance.