Purpose: High resolution imaging of corneal epithelial cells can be achieved non-invasively with in vivo confocal microscopy (IVCM). Currently no automated method for quantitative epithelial cell analysis exists; manual methods are tedious and time-consuming. Herein, we describe an automated, deep-learning approach to epithelial cell density quantification.

Methods: Different machine learning approaches were explored resulting in a dual-stage approach to performing epithelial cell counting. The first stage implements a K-Nearest Neighbors (KNN) model to assess the “countability” of cells in a corneal image while a second stage implements a Random Forest model to predict cell counts for the images identified as “countable.” All models were trained and tuned on a dataset of 212 anonymized IVCM (HRT 3 RCM, Heidelberg Engineering GmbH, Heidelberg, Germany) epithelial wing and basal cell images labeled with values generated by trained human graders using manual cell counting software. Results for the dual-stage approach were evaluated against manual results.

Results: A dual-stage automated approach utilizing a KNN Classification model paired with a Random Forest counting model was the highest performing of the explored approaches. The KNN Classification model used to assess the “countability” of an image achieved an accuracy classification score of 79.1% while the Random Forest counting model used to predict epithelial cell counts had a mean absolute percentage error of 8.0% and a corresponding mean error of 10.88 cells per mm^2. Single model approaches achieved worse results with individual KNN Classification and Random Forest models with mean errors of 82.72 cells per mm^2 and 61.60 cells per mm^2, respectively.

Conclusions: A dual-stage approach resulted in significant improvement in epithelial cell density quantification compared to single model approaches. Compared to manual and automated single model approaches, the dual-stage automated approach improves grading speed while maintaining good accuracy with the potential for further improvement with additional study.