Yedidya, Tamir
Description
Dry Eye Syndrome is a common disease in the western world, with effects from uncomfortable itchiness to permanent damage to the ocular surface. Almost 5 million Americans over 50 years old suffer from dry eye. A conservative estimate shows that approximately 17 million Americans have contact lens related dry eye -one of the main factors to contact lens discontinuation. In addition, the incidence of the disease is on the rise. Nevertheless, there is still no gold standard test that can reliably...[Show more] detect dry eye. One of the most commonly used tests by clinicians to detect dry eye is the Fluorescein Break Up Time (FBUT). However, results vary a lot between clinicians. Other tests such as observing the tear meniscus height are also performed regularly by the clinicians but not necessarily in conjunction with the FBUT test. Therefore there is a real need for a reliable, robust and operator-dependent method to evaluate dry eye. To our knowledge, no previous research has been conducted on automatic evaluation of dry eye in fluorescein images. In this thesis, we present new algorithms to automatically detect various dryness signs and make a number of original contributions. The first problem we address is how to detect the dry areas in fluorescein videos of the anterior of the eye, which are captured using a portable camera. We present a new multi-step algorithm which first locates the iris in each image in the video, then aligns the images according to the location of the iris and finally analyzes the aligned video to find the regions of dryness. We produce a novel segmentation result called dryness image, which depicts the various degrees of tear film thinning over the corneal surface. Then, we demonstrate through experiments that there is a large variation in the estimated Break Up Time (BUT) between clinicians and no ground-truth can be defined. To overcome that, we define a new value based on the clinical definitions of the BUT. These definitions are converted to image processing properties and an estimate of the BUT is computed using temporal analysis of the aligned video. We demonstrate that our new value is in the accepted range of the BUT values provided by the clinicians. We present an extension to the dryness algorithm, which is based on transforming the video to a volume by considering each video frame as a slice in a 3D volume. On a volume, a temporal monotonic constraint can be applied between pixels in consecutive slices. The constraint enforces the clinical definition of tear film thinning over time -the amount of fluid cannot increase while not blinking. The constraint is applied directly into the cost function and the whole volume is segmented simultaneously using graph-cuts. As a consequence, the approach is more robust and less sensitive to alignment errors. Finally, we generalize the idea and explain how monotonic constraints can be applied to other imaging modalities. In the last part of the thesis, we develop a new algorithm to evaluate the tear meniscus height and shape using graph-cuts. We formulate the segmentation problem using asymmetric cost functions and demonstrate its power and usefulness for the task. The asymmetry induces which directional moves are permitted in the minimization process and thus produces a result that adheres to the known shape properties of the tear meniscus. The iterative algorithm provides simultaneously the best segmentation result and shape prior of the meniscus.
Items in Open Research are protected by copyright, with all rights reserved, unless otherwise indicated.