Abstract:
The cornea typically constitutes two-thirds of the eye’s optical power; a healthy one has a dome shape. However, a vision disorder so-called Keratoconus, which is especially prevalent in the Middle East and Gulf area, may result in a cone shape of the cornea, leading to significant loss of visual acuity. As a healing procedure, ophthalmologists use a technique so-called corneal cross-linking to halt the progression of Keratoconus. One indicator of corneal cross-linking success is arguably the presence and depth of the stromal demarcation line. In addition, corneal haze beyond the demarcation line can be an ominous sign of loss of corneal transparency after cross-linking, which is a much-dreaded side-effect of the procedure. To date, to observe corneal haze, ophthalmologists use slit lamps and-or they observe Optical Coherence Tomography (OCT) micrometer resolution images of the corneal tissue. These techniques are subjective, time consuming and error-prone. In this thesis, we propose a novel technology to automatically detect and measure corneal haze and demarcation line in OCT images. To achieve so, we propose new image analysis algorithms and make use of existing libraries such as OpenCV; additionally, we propose a customized machine learning approach used for OCT image classification. The new automated tool provides the user with haze statistics as well as visual annotation reflecting the shape and location of the haze and demarcation line in the cornea; it also detects subtle changes in the cornea over time. Our experimental results and analysis demonstrate the efficacy and effectiveness of our new tool in accurately detecting and measuring the demarcation line depth, in comparison to manual measurements, in a much faster manner. The Intraclass correlation coefficients (ICC) and Pearson correlation coefficients (PCC) between the automated detection software and two human operators are measured as 0.945 and 0.951, and 0.910 and 0.918, respectively. The ICC for inter-operators reproducibility is 0.882, a
Description:
Thesis. M.S. American University of Beirut. Department of Computer Science, 2017. T:6639
Advisor : Dr. Ahmad R. Dhaini, Assistant Professor, Computer Science ; Committee members : Dr. George Turkiyyah, Professor, Computer Science ; Dr. Shady Elbassuoni, Assistant Professor, Computer Science.
Includes bibliographical references (leaves 80-84)