Methods for automated analysis of macular OCT data

Optical coherence tomography (OCT) is fast becoming one of the most important modalities for imaging the eye. It provides high resolution, cross-sectional images of the retina in three dimensions, distinctly showing its many layers. These layers are critical for normal eye function, and vision loss may occur when they are altered by disease. Specifically, the thickness of individual layers can change over time, thereby making the ability to accurately measure these thicknesses an important part of learning about how different diseases affect the eye. Since manual segmentation of the layers in OCT data is time consuming and tedious, automated methods are necessary to extract layer thicknesses. While a standard set of tools exist on the scanners to automatically segment the retina, the output is often limited, providing measurements restricted to only a few layers. Analysis of longitudinal data is also limited, with scans from the same subject often processed independently and registered using only a single landmark at the fovea. Quantification of other changes in the retina, including the accumulation of fluid, are also generally unavailable using the built-in software. In this thesis, we present four contributions for automatically processing OCT data, specifically for data acquired from the macular region of the retina. First, we present a layer segmentation algorithm to robustly segment the eight visible layers of the retina. Our approach combines the use of a random forest (RF) classifier, which produces boundary probabilities, with a boundary refinement algorithm to find surfaces maximizing the RF probabilities. Second, we present a pair of methods for processing longitudinal data from individual subjects: one combining registration and motion correction, and one for simultaneously segmenting the layers across all scans. Third, we develop a method for segmentation of microcystic macular edema, which appear as small, fluid-filled, cystoid spaces within the retina. Our approach again uses an RF classifier to produce a robust segmentation. Finally, we present the development of macular flatspace (MFS), a computational domain used to put data from different subjects in a common coordinate system where each layer appears flat, thereby simplifying any automated processing. We present two applications of MFS: inhomogeneity correction to normalize the intensities within each layer, and layer segmentation by adapting and simplifying a graph formulation used previously