A software system for automatic estimation of corneal epithelial cell density in confocal microscopy

Purpose: The problem of reliable automatic estimation of density of corneal epithelial cells in images from confocal microscopy was addressed. The reliability of estimated densities should be comparable to that of manual cell count. Methods: The spatial frequencies contained in epithelial images can be extracted with suitable mathematical techniques (2-dimension Discrete Fourier Transform, DFT). An algorithm for reliably identifying the spatial frequency information and for deriving from it an estimate of the cell density has been developed. A preliminary entropy-based pre-processing was carried out to extract from the whole image the ROI (region of interest) where cells are visible. A prototype of the whole algorithm was implemented in the Matlab® language and run on a personal computer. A preliminary evaluation was performed on a data set containing 24 images from normal subjects (see e.g. enclosed image), acquired with a ConfoScan 3 confocal microscope (Nidek Technologies, Italy). Reference manual counts were performed on each image by an experienced optometrist. Results: Mean percent difference of automatic densities vs. manual ones was ?1%, with std dev 10% and range [?17% ? +18%]. Mean percent absolute difference was 9%, with std dev 4% and range [+1% ? +18%]. Running times of the prototype were in the order of 40 seconds per image. Conclusions: A new algorithm was developed for the automatic estimation of epithelial cell density. A preliminary evaluation of the proposed technique confirmed its capability of reliably estimating corneal epithelial cell density in confocal images of normal subjects. Implementation of the algorithm with a more efficient computer language, e.g. C++, will allow execution times in the order of 2-3 seconds.