Automated tri-image analysis of stored corneal endothelium

Abstract

Background: Endothelial examination of organ culture stored corneas is usually done manually and on several mosaic zones. Some banks use an image analyser that takes account of only one zone. This method is restricted by image quality, and may be inaccurate if endothelial cell density (ECD) within the mosaic is not homogeneous. The authors have developed an analyser that has tools for automatic error detection and correction, and can measure ECD and perform morphometry on multiple zones of three images of the endothelial mosaic.

Methods: 60 human corneas were divided into two equal groups: group 1 with homogeneous mosaics, group 2 with heterogeneous ones. Three standard microscopy video images of the endothelium, graded by quality, were analysed either in isolation (so called mono-image analysis) or simultaneously (so called tri-image analysis), with 50 or 300 endothelial cells (ECs) counted. The automated analysis was compared with the manual analysis, which concerned 10 non-adjacent zones and about 300 cells. For each analysis method, failures and durations were studied according to image quality.

Results: All corneas were able to undergo analysis, in about 2 or 7.5 minutes for 50 and 300 ECs respectively. The tri-image analysis did not increase analysis time and never failed, even with mediocre images. The tri-image analysis of 300 ECs was always most highly correlated with the manual count, particularly in the heterogeneous cornea group (r=0.94, p<0.001) and prevented serious count errors.

Conclusions: This analyser allows reliable and rapid analysis of ECD, even for heterogeneous endothelia mosaics and mediocre images.

Figure 1

Different grades of standard microscopy video image of the endothelium. These very wide field images (750 000 μm²) were archived in TIFF format. (A) Adequate quality image with poorly delineated cell contours, mediocre contrast, large blocks of invisible cells. (B) Medium quality image with fairly well delineated cell contours but high background noise and large blocks of invisible cells. (C) Good quality image with well delineated cell contours, no background noise, cells visible on practically all the image except one fold zone (arrow). (D) Automated error detection tool. The contours of cells exceeding 1500 μm² or the form of which is unrealistic (longest length greater than double the shortest width) appear violet to alert the technician. He could then either validate the cells if he deemed them acceptable, or touch them up by adding or erasing certain segments. In this example, six cells were genuinely large in area (black arrows), the others needed touching up (red arrows). Scale bar 200 μm

Figure 2

Final result of analyses in mono-image (A) and tri-image (B) modes, sent by the cornea bank to the surgeon. This document, saved in HTML format, can be remotely transmitted. It contained all quantitative data (ECD, cell count) and morphometric data (area variation coefficient, min/max/mean/standard deviation of cell area, histogram of relative frequency distribution of cell areas, percentage of hexagonality). The one or three non-analysed and analysed video images of the endothelium are also printed, as is a non-specified image (in our case, an example of a graft). The technician also had a text box to note any comments.

Figure 3

The relations between each of the eight automated analysis methods and manual count for homogeneous corneas (group 1, n=30). Whatever the analysis method, discrepancies with the manual count were rare. Correlations were best with the tri-image mode, for 300 and even 50 cells. For all correlations, p was <0.001.

Figure 4

The relations between each of the eight automated analysis methods and manual count for heterogeneous corneas (group 1, n=30) The mono-image 50 or 300 counts often gave rise to serious discrepancies. Correlations were best (r=0.94) in mono-image 300 and tri-image 50 modes (r=0.91). For all correlations, p was <0.001.