Image Analysis for the Assessment of Pathologies in Retinal Tissue
thesisposted on 2014-02-24, 00:00 authored by Fatimah Mohammad
Photoreceptor cell degeneration due to disease may lead to loss of visual acuity. With the recent availability of adaptive optics (AO) and spectral domain optical coherence tomography (SD-OCT) imaging systems, it is now possible to image the retina at the cellular level. Many of the challenges involved with retinal image analysis are due to the low contrast and system noise observable in the images. Photoreceptor cell regularity and the integrity of the photoreceptor segment junction layer are studied as the photoreceptor cells play a fundamental role in human vision. The goal of this research is to develop automated methods which can provide qualitative and quantitative metrics that aim to describe the state of healthy or diseased retinas. We present a novel framework which can potentially link the observations made in AO retinal images with those made in SD-OCT en face retinal images. In our preliminary work, we focus on photoreceptor cell regularity estimation by the quantification of cells in AO retinal images of the human eye. The issues addressed include low contrast images as well as varying cellular packing arrangement within a single image. We initially developed a content-adaptive filtering method which analyzes image frequency content in the spatial domain using intensity profiles of the image to identify photoreceptor cells. To account for false positive detection due to noise, we describe an image model in the frequency domain using a windowed, two-dimensional (2D) lattice of pulses which represent the cells and characterize the frequency content as decaying frequency domain pulses on the reciprocal lattice. This model uses a small-extent, block-based, 2D Discrete Fourier transform (DFT) to obtain the parameters of an adaptive, circularly-symmetric band-pass filter that is applied to the image to extract the underlying cellular structure and remove high and low frequency contamination. Our automated results of cell detection compared well with manual results on computer-generated test and retinal images. There is a strong association between the integrity of the ISOS layer and the photoreceptor health. We therefore developed a method for pathology segmentation in fluorescein angiograms as test images, and en face retinal images of the ISOS layer of patients with age-related macular degeneration and diabetic retinopathy. We develop a level-set method based on the classical Chan-Vese model and exploit a priori knowledge of the shape and intensity distribution, allowing the use of projection profiles to detect the presence of pathologies. Our method provides improved speed and reliability in the segmentation which may fail in classical algorithms with an incorrect choice of initial contour. It was of interest to analyze all retinal cell layers in an en face manner, to investigate if pathologies in an overlying retinal layer could cause artifacts or shadows to appear on the ISOS layer. In this case, the pathology that appears to be on the ISOS layer would be due to optical factors. If however the pathology on the ISOS layer is unrelated to overlying pathology, it would be indicative of neural dysfunction. We develop a level-set method that incorporates shape priors, defined by Fourier descriptors, to guide the evolution of the curve towards objects matching a target shape. We apply our method to en face images representing seven layers of the retina. The level-set function is defined such that it evolves across the layers and adapts to the pathologies present in the image. In order to determine the origin of pathology, we measure the co-localization across overlapping layers. We show how our method overcomes the common problems encountered by other proposed level-set models by comparing our method to a well-known distance-regularized method which does not use shape priors. The comparison clearly shows how use of our method is more effective due to the incorporation of Fourier-based shape priors.