Image analytics for prediction of keratoplasty failure Summary We will create specialized image analytics software for prediction of keratoplasty (penetrating, endothelial) fail- ure from specular-reflection corneal endothelial cell (EC) images. Keratoplasties are the most common tissue transplant, with roughly a 10% failure rate, leading to blindness, patient discomfort/anxiety, and repeat kerato- plasties with a higher chance for failure than the initial procedure. With successful predictive image analytics, we will be in a position to identify transplanted corneas at risk and possibly treat them more aggressively with topical corticosteroids or other measures to prevent failure. Since a functional endothelial cell (EC) layer is necessary for the active ionic-pump-driven redistribution of fluid necessary to maintain the clear cornea, EC images have been analyzed as an indicator of cornea health. The normal EC layer exhibits high cell density arranged in a predominantly regular, hexagonal array. We will build on the use of existing quantitative bi- omarkers from EC images (EC density, coefficient of variation of cell areas, and hexagonality) used to evaluate cornea health. We will compute additional image features associated with local and long-range cell disarray, image attributes relevant to keratoplasty rejection, and traditional features from computer vision. Including this combination of features will provide rich inputs to machine-learning classifiers aimed at predicting future out- comes (e.g., failure or no failure). We will apply methods to a large aggregation of well-curated data from pre- vious NIH-funded studies at Case Western Reserve University (CWRU) and from previous studies at the Neth- erlands Institute for Innovative Ocular Surgery (NIIOS). Our team consists of image processing experts, oph- thalmologists, and staff from the CWRU Department of Ophthalmology and Visual Sciences and University Hospitals (UH) Eye Institute’s Cornea Image Analysis Reading Center (CIARC), which is well-known for rigor- ous, highly repeatable assessment of conventional quantitative biomarkers in a large number of multi- institutional clinical trials. Together, our goal will be to determine if this “second generation” analysis of EC im- ages can lead to prediction of keratoplasty failure. If successful, this project will lead to software which can be translated to support research and clinical practice.

Narrative Our goal is to create image analytic software that will predict the risk of keratoplasty failure from readily ob- tained corneal endothelial cell images. With knowledge of eyes at risk, physicians will be able to tailor treat- ments to improve cornea transplant success, thereby very positively impacting patients’ health.