Segmentation of Corneal Endothelial Cells Contour by Means of a Genetic Algorithm

Abstract

Corneal images acquired by in-vivo microscopy provide clinical in- formation on the cornea endothelium health state. The reliable estimation of the clinical morphometric parameters requires the accurate detection of cell con- tours in a large number of cells. Thus for the practical application of this analy- sis in clinical settings an automated method is needed. We propose the automat- ic segmentation of corneal endothelial cells contour through an innovative technique based on a genetic algorithm, which combines information about the typical regularity of endothelial cells shape with the pixels intensity of the actu- al image. Ground truth values for the clinical parameters were obtained from manually drawn cell contours. Results show that an accurate automatic estima- tion is achieved: for each parameter, the mean difference between its manual estimation and the automated one is always less than 4%, and the maximum difference is always less than 7%. The normal human corneal endothelium is a single layer of uniformly sized cells with a predominantly hexagonal shape, but this regular tessellation is affected by age and pathologies (1). Thus, the analysis of the main morphometric parameters of corneal endothelium provides clinical information capable to describe the cornea health state. Namely, endothelial cell density (ECD), polymegethism (differences in cell size ex- pressed as fractional standard deviation of cell areas), and pleomorphism or hexago- nality coefficient (fraction of hexagonal cells over the total number of cells) are commonly used as parameters to quantitatively characterize the endothelial cells' condition. In order to make this analysis practical in clinical settings, a computerized method that fully automates the segmentation procedure would be needed. The fundamental problem with automated endothelial analysis is to correctly identify the cells' contour, a necessary prerequisite for the estimation of the clinical parameters (2). The automat- ic recognition of the cell boundaries is a challenging task because of the noise that make the contour difficult to be distinguished even by an expert, and the substantial differences between an image and the other as regards size and appearance of the