Automated identification and quantification of activated dendritic cells in central cornea using artificial intelligence.

Purpose: To validate an algorithm quantifying activated dendritic cells (aDCs) using in-vivo confocal microscopy (IVCM) images.
Methods: IVCM images obtained at the Miami Veterans Affairs Hospital were retrospectively analyzed. ADCs were quantified both with an automated algorithm and manually. Intra-class-correlation (ICC) and a Bland-Altman plot were used to compare automated and manual counts. As a secondary analysis, individuals were grouped by Dry Eye (DE) subtype: 1) aqueous-tear deficiency (ATD; Schirmer's test ≤5 mm); 2) evaporative DE (EDE; TBUT≤5s); or 3) control (Schirmer's test>5 mm; TBUT>5s) and ICCs were re-examined.
Results: 173 non-overlapping images from 86 individuals were included in this study. The mean age was 55.2 ± 16.7 years; 77.9% were male; 20 had ATD; 18 EDE and 37 were controls. The mean number of aDCs in the central cornea quantified automatically was 0.83 ± 1.33 cells/image and manually was 1.03 ± 1.65 cells/image. A total of 143 aDCs were identified by the automated algorithm and 178 aDCs were identified manually. While a Bland-Altman plot indicated a small difference between the two methods (0.19, p < 0.01), the ICC of 0.80 (p = 0.01) demonstrated excellent agreement. Secondarily, similar results were found by DE type with an ICC of 0.75 (p = 0.01) for the ATD group, 0.80 (p = 0.01) for EDE, and 0.82 (p = 0.01) for controls.
Conclusions: Quantification of aDCs within the central cornea may be successfully estimated using an automated machine learning based algorithm. While this study suggests that analysis using artificial intelligence has comparable results with manual quantification, further longitudinal research to validate our findings in more diverse populations may be warranted.
(Published by Elsevier Inc.)