Quantifying corneal immune cells from human in vivo confocal microscopy images: Can manual quantification be improved with observer training?

Abstract

Manually quantifying immune cells (ICs), commonly considered dendritic cells, in the corneal epithelium from in vivo confocal microscopy (IVCM) images can be influenced by observer bias. This study sought to evaluate the repeatability of manual IC quantification. Cell counts were first performed for 184 non-overlapping IVCM images by a single observer. Quantifications were undertaken to establish the total cell numbers per image, and the numbers of three cell morphological subtypes: mature ICs (with elongated dendrites), immature ICs (with short- or non-discernible dendrites) and globular cells (with large bodies and no visible dendrites). Cell counts were then repeated by the same observer, and independently undertaken by a second observer. Prior to these counts, both observers undertook an agreement ‘training’ process to define IC appearance and delineate the morphological subtypes.Total IC counts demonstrated excellent intra- and inter-observer reliability (intraclass correlation coefficients (ICC) > 0.90). Bland-Altman plots showed that interobserver measurement bias increased as a function of the total IC number in the image prior to consensus training. For total IC counts after the observer training process, there was no significant interobserver measurement bias. For IC morphological subtypes, there was a positive relationship between the mean inter-observer difference and average cell count for mature ICs and globular cells, but not immature ICs.In conclusion, higher variability in manual corneal IC counts exists when more cells are present in an IVCM image. Implementing an observer training process reduced inter-observer variability and minimised systematic measurement error.