Abstract
Purpose :
Activated dendritic cells (aDCs) have recently been identified as potential biomarkers indicating the presence of a systemic auto-immune disease in individuals with dry eye (DE). However, their evaluation with in-vivo confocal microscopy (IVCM) is relatively subjective. Therefore, there is a need for a standardized identification method of aDCs to improve generalizability. Our aim was to validate an algorithm that automatically identifies and quantifies aDCs using IVCM images of the central cornea.
Methods :
A retrospective analysis of IVCM images obtained from individuals seen at the eye clinic at the Miami Veterans Affairs Hospital was performed. Images from individuals with corneal scarring were excluded due to potential confounding. Artificial intelligence was utilized through the use of an automated aDC counter which was developed using transfer learning with IVCM images. IVCM images used in the development of this algorithm were acquired prior to the start date of this study and did not include any of the same patients. ADCs were manually quantified based on morphology by reviewers that were masked to the algorithm findings, and intra-class correlation (ICC) was used to compare automated and manual counts. Algorithm accuracy was defined as aDC counts within 1 cell compared to manual quantification.
Results :
A total of 193 non-overlapping IVCM images from 110 individuals were included. The mean age of individuals included in the study was 55.5±18.4 years; 70.0% were male, 58.2% self-identified as White and 24.5% as Hispanic. There was no statistically significant difference in the mean aDC count between automated and manual quantifications (1.06±1.82 cells/image vs 1.21±1.98 cells/image, p=0.47). The algorithm identified a total of 207 aDCs within the dataset, out of which 184 were manually verified as aDCs. The automated algorithm performed the aDCs count with 87% accuracy and an ICC of 83% (p<0.01).
Conclusions :
The number of aDCs in the central cornea can be successfully estimated with the use of artificial intelligence with comparable results to manual quantification. Further assessment is needed before the widespread clinical implementation of an automated aDC counter.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.