Purchase this article with an account.
Alfredo Ruggeri, Enea Poletti; AUTOMATIC ESTIMATION OF MORPHOMETRIC PARAMETERS IN CORNEAL ENDOTHELIUM IMAGES. Invest. Ophthalmol. Vis. Sci. 2014;55(13):2070.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To provide ophthalmologist with a fully automated algorithm for the reliable identification of the endothelial cell contours. From the contour detection in a fairly large number of cells, a reliable estimation of the main morphometric parameters (endothelial cell density (ECD), pleomorphism, polymegethism) can then be derived.
Thirty images of corneal endothelium were acquired with a specular endothelial microscope (SP-3000P, Topcon Co., Japan) from both healthy and pathological subjects. The region of interest covered an area of about 0.1 mm2, including an average of 220 cells. Ground-truth reference values were obtained by estimating the morphometric parameters from manually drawn cell contours. The proposed method analyzes a fixed-size neighborhood of each image pixel and computes three key features (signatures) of the pixel, roughly expressing its probability of belonging to each of the geometrical structures that define a cell: vertex, side, and body. Multi-scale 2-dimensional matched filters were specifically designed to the detect each signature value. The different values they assume in each image pixel were then used as input features for a Support Vector Machine classifier, used to classify that pixel as being a vertex, side, or body pixel. Combining all the pixels classified as vertex or side eventually provided the identification of the cells contours, from which morphometric parameters could easily be computed.
The average percent differences from ground-truth reference of the estimated morphometric parameters in the 30 images were 1.14% for ECD, -4.41% for pleomorphism and -13.40% for polymegethism. The processing time was 5~10 s per image.
The estimates of the corneal morphometric parameters provided by the proposed, fully automatic algorithm are in good agreement with ground truth values, obtained with a careful, time-consuming manual analysis. Our system can analyze regions containing hundreds of cells and thus provide a very high accuracy for the estimated parameters. Thanks to the short run time, this could be further increased by performing in less than one minute the analysis on several images per subject.
This PDF is available to Subscribers Only