Impact of image selection on corneal nerve and dendritic cell density quantification using in vivo confocal microscopy

Abstract

Aim

The aim of this study was to measure the impact of image selection on quantification of corneal nerve fibre length density (CNFL) and dendritic cell density (DCD) using in-vivo confocal microscopy images.

Methods

We used images from 80 participants (mean age 51 years; 55 female), averaging 1030 images per participant. CNFL and DCD were quantified using independently selected sets and same-image sets. Inter-grader agreement was computed using 10 or 5 images per participant (patient-level data point). Grader agreement was quantified by intraclass correlation coefficients (ICC).

Results

Using independently selected images, the ICC for CNFL was 0.82 with 10 images, which was significantly higher by 0.07 than with 5 images (p = 0.009). For DCD, the ICC was 0.92 with 10 images and was not significantly higher than with 5 images. Using same-image sets, the ICC for CNFL was 0.94 with both 5 and 10 images. For DCD, the ICC improved by 0.03 when 10 images were used compared with 5 images (p = 0.02).

Conclusion

Independent-image selection gives poorer reproducibility of CNFL and DCD than multiple graders using the same-image sets. The same-image sets yielded excellent agreement, suggesting that studies relying on a single selection of five or fewer images should be interpreted with caution, as a single selection is more prone to rater bias. Because a single set of selected images reflects real-world practice, we recommend that future studies use multiple graders, each independently selecting and grading images, with more than five images per participant where feasible.