Abstract
Purpose :
We aim to develop a deep learning (DL) network capable of delineating corneal endothelial cells (CEC) from specular microscopy (SM) images, and to develop algorithms for counting and morphology identification of CEC.
Methods :
A set of 600 specular microscopy (SM) images were acquired using CEM-530, Nidek, Japan at baseline, month 1, and month 3 from 10 subjects, employing scanning at 15 distinct fixation positions on the cornea (1 central position, 8 paracentral positions at a diameter of 1.3 mm, and 6 peripheral positions at a diameter of 7.3 mm). Initially, 100 SM images from healthy corneas underwent visibility enhancement and manual segmentation, with delineation of corneal endothelial cell (CEC) boundaries. A supervised DL network was trained to identify CEC, and accuracy was measured using the Dice coefficient (DC). The network's robustness was subsequently assessed using SM images from unhealthy corneas as an independent test set. Additionally, an algorithm was developed for the quantitative assessment of CEC, including cell count, percentage of hexagonal cells, coefficient of variation, and other morphological features. These features were longitudinally tracked at all 15 positions, and visual representation using color pi-chart was employed for comprehensive corneal health assessment.
Results :
Our segmentation network demonstrated high precision in delineating CEC from SM images, achieving an DC of 0.89±0.02. The DL network, initially trained on SM images from healthy corneas, exhibited robust performance when applied to images from unhealthy corneas, yielding a DC of 0.86±0.02. Notably, longitudinal cell counting across 15 different positions revealed variations in cell numbers, indicating dynamic changes in cell populations over the course of progression.
Conclusions :
The DL network efficiently differentiated CEC and executed segmentation tasks with remarkable precision. The longitudinal investigation across 15 distinct fixation positions successfully revealed the progression patterns within each subject. Our methodology shows potential in differentiating unhealthy corneas from healthy ones by tracking morphological data over time.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.