Assessing abnormal corneal endothelial cells from in vivo confocal microscopy images using a fully automated deep learning system

Abstract

Background: The goal of this study is to develop a fully automated segmentation and morphometric parameter estimation system for assessing abnormal corneal endothelial cells (CECs) from LASER in vivo confocal microscopy (IVCM) images. Methods: First, we developed a fully automated deep learning system for assessing abnormal CECs using a previous development set composed of normal images and a newly constructed development set composed of abnormal images. Second, two testing sets, one with 169 normal images and the other with 211 abnormal images, were used to evaluate the clinical validity and effectiveness of the proposed system on LASER IVCM images with different corneal endothelial conditions, particularly on abnormal images. Third, the automatically calculated endothelial cell density (ECD) and the manually calculated ECD were compared using both the previous and proposed systems. Results: The automated morphometric parameter estimations of the average number of cells, ECD, coefficient of variation in cell area and percentage of hexagonal cells were 257 cells, 2648 ± 511 cells/mm2, 32.18 ± 6.70% and 56.23 ± 8.69% for the normal CEC testing set and 83 cells, 1450 ± 656 cells/mm2, 34.87 ± 10.53% and 42.55 ± 20.64% for the abnormal CEC testing set. Furthermore, for the abnormal CEC testing set, Pearson’s correlation coefficient between the automatically and manually calculated ECDs was 0.9447; the 95% limits of agreement between the manually and automatically calculated ECDs were between 329.0 and − 579.5 (concordance correlation coefficient = 0.93). Conclusions: This is the first report to count and analyze the morphology of abnormal CECs in LASER IVCM images using deep learning. Deep learning produces highly objective evaluation indicators for LASER IVCM corneal endothelium images and greatly expands the range of applications for LASER IVCM.