A reproducible automated segmentation algorithm for corneal epithelium cell images from in vivo laser scanning confocal microscopy

Abstract

Purpose: To evaluate an automated process to find borders of corneal basal epithelial cells in pictures obtained from in vivo laser scanning confocal microscopy (Heidelberg Retina Tomograph III with Rostock corneal module).

Methods: On a sample of 20 normal corneal epithelial pictures, images were segmented through an automated four-step segmentation algorithm. Steps of the algorithm included noise reduction through a fast Fourier transform (FFT) band-pass filter, image binarization with a mean value threshold, watershed segmentation algorithm on distance map to separate fused cells and Voronoi diagram segmentation algorithm (which gives a final mask of cell borders). Cells were then automatically counted using this border mask. On the original image either with contrast enhancement or noise reduction, cells were manually counted by a trained operator.

Results: The average cell density was 7722.5 cells/mm(2) as assessed by automated analysis and 7732.5 cells/mm(2) as assessed by manual analysis (p = 0.93). Correlation between automated and manual analysis was strong (r = 0.974 [0.934-0.990], p < 0.001). Bland-Altman method gives a mean difference in density of 10 cells/mm(2) and a limits of agreement ranging from -971 to +991 cells/mm(2) . Visually, the algorithm correctly found almost all borders.

Conclusion: This automated segmentation algorithm is worth for assessing corneal epithelial basal cell density and morphometry. This procedure is fully reproducible, with no operator-induced variability.

Keywords: In vivo laser scanning confocal microscopy; cell morphometry; corneal epithelium; image processing.