Cornea:

Automatic Nerve Tracing

3D Reconstruction of the Cornea

Automatic Endothelial Cell Count

Identification of Endothelial Cells


Automatic Endothelial Cell Count

A. Ruggeri, E. Grisan, J. Schroeter. Evaluation of repeatability for the automatic estimation of endothelial cell density in donor corneas. Br J Ophthalmol 91:1213-1215, Sep 2007.
A. Ruggeri, E. Grisan, J. Jaroszewski. A new system for the automatic estimation of endothelial cell density in donor corneas. Br J Ophthalmol 89:306-311, Mar 2005.
M. Foracchia, A. Ruggeri. Automatic estimation of endothelium cell density in donor corneas by means of Fourier analysis. Med Biol Eng Comput 42(5): 725-731, Sep 2004.

Optical microscopy images of corneal endothelium are commonly acquired and analyzed for the clinical assessment of cornea quality, particularly for diagnostic purposes or for evaluating the possible outcome of cornea transplantation. When the images are acquired with a typical microscope, they usually cover a wide area of the specimen, containing a large number of cells (5000-10000). Despite of this, manual cell counts are routinely performed at Eye Banks to determine the density of endothelium cells.
To develop an automatic system to perform this task, the application of the conventional edge detection techniques or of the morphological operation tecniques, would fail because of the actual lack of visible edges. Also, the information required by clinicians is usually just that of the cell density value in the analyzed specimen, and a detailed morphological analysis is not required.

Two approaches to the estimation problem have been devised. One is based on Fourier analysis, the other on the estimation of the generator matrix of the cellular lattice.



Fourier Analysis

We addressed the problem by analyzing the image in the frequency domain, since the density of the cells pattern can be related to the spatial frequency of the cells.




Original sample endothelial image, obtained from a optical microscope with a 100x magnification.




On the left: analysis of the frequency domain: raw data (top panel), smoothed data (middle panel), derivative of the smoothed spline (bottom panel). On the rigth: estimated cell spatial frequency and confidence interval.





Pass band filtered version of the original image, with central frequency the estimated cell spatial freqency and band-width the confidence interval. It is evident that the information retained is the cell pattern.




return to top



Lattice Estimation

The analysis of microscopy images of corneal endothelium is routinely carried on at eye banks to assess cornea health state and quality. Ideal endothelia are made of regular hexagonal cells of similar area, but this regular tessellation is affected by age and pathologies. A quantitative analysis may provide information on cell density, polymegethism (distribution of cell sizes), or pleomorphism (distribution of number of cell sides).
This information could be easily extracted if the cell contours are identified in the image, and several prototype systems for the automatic analysis of corneal endothelium based on the identification of cell contours have been proposed. Unfortunately, due to cornea anatomy and specific features of image acquisition, images are often blurred and noisy, so that contour recognition is rather difficult and these systems often require operator interaction to correct errors: none of the proposed systems have been able to obtain a reliable estimation of quantitative indexes without tedious and time-consuming manual editing.

Recently, an automatic system based on the Fourier analysis has been proposed, which does not require the difficult identification of cell contours and estimates the cell density from the information about the spatial cell frequency. This system, however, is well suited for images with a high number of cells (typically those acquired with a 100x magnification), whereas it is easily confounded in images with higher magnification, where the cell frequency has the same order of magnitude of the low frequencies due to illumination, corneal folds, or out-of-focus areas.

We propose here a new method, still based on the regular tessellation feature of endothelium cells, which estimates the generator matrix of the cellular lattice, from which the density is easily obtained. Since cells have different spatial orientations throughout the corneal image, the base matrix may vary substantially from one region to another. A local estimation is therefore performed, in order to minimize the effects of this orientation variability. This also provides an additional result an estimate of the variability of the corneal density (polymegethism).





False colour map of cell density ranging from 300 cells/mm2 (blue) to
6000 cells/mm2 (red), superimposed to the original image.


return to top