ESCRS - PP09.01 - Analysis Of Cornea Endothelial Imaging Using Deep Learning In Fuchs Dystrophy

Analysis Of Cornea Endothelial Imaging Using Deep Learning In Fuchs Dystrophy

Published 2025 - 43rd Congress of the ESCRS

Reference: PP09.01 | Type: Poster | DOI: 10.82333/d9ae-re14

Authors: Kai Yuan Tey* 1 , Satish Panda 2 , Qiu Ying Wong 3 , Ezekiel Ze Ken Cheong 4 , Marcus Ang 5

1Singapore National Eye Centre,Singapore,Singapore, 22School of Mechanical Science,Indian Institute of Technology Bhubaneswar,Odisha,India, 3Singapore Eye Research Institute,Singapore,Singapore, 4Duke-NUS Medical School,Singapore,Singapore, 5Cornea and External Eye Disease ,Singapore National Eye Centre,Singapore,Singapore;Singapore Eye Research Institute,Singapore,Singapore

Purpose

To evaluate the use of a deep learning network (DLN) in the analysis of widefield specular microscopy (WFSM) images in eyes with Fuchs’ endothelial dystrophy (FECD).

Setting

A cross-sectional, clinical observational study in a tertiary care setting.

Methods

A total of 100 FECD eyes underwent WFSM (CEM-530, Nidek Co. Ltd, Japan). Images were graded using a standardized quality score. Central region images were compared to paracentral and peripheral regions based on image quality, and morphometric parameters: endothelial cell density (ECD), coefficient of variation (CV), and hexagonality (HEX). A deep learning network (DLN) using U-Net architecture was developed, and trained on specular microscopy (SM) images from 50 FECD and 50 control eyes (70% training, 30% validation). It was tested on an external longitudinal dataset (baseline and 1-month). Comparison of DLN performance and in-built manual analysis was made using Sørensen–Dice coefficient, and paired t-tests (for morphometric parameters).

Results

A good intergrader agreement was observed for both SM image quality (κ= 0.967, 95% CI: 0.959–0.976), and FECD disease severity grading (κ= 0.891, 95% CI: 0.786–0.995). There were no significant differences between paracentral/peripheral ECD in eyes without edema (p < 0.001–0.003), and with subclinical edema (p < 0.001–0.011). Overall, DLN-driven segmentation was comparable to manual segmentation (dice coefficient = 0.85 ± 0.063). Comparing the DLN-derived ECD against manual analysis, the mean ECD in the central region was comparable between the two methods (ECDDLN = 2715.10 ± 185.21 cells/mm2 vs ECDManual = 2507.80 ± 271.36 cells/mm2, p = 0.196).

Conclusions

We proposed a novel application of deep learning for the analysis of widefield corneal endothelial imaging. Combined with the integration of a progression visualization tool, this approach allows efficient auto-analysis and organization of the large datasets generated by wide-field imaging, and helps clinicians interpret results more effectively and quickly, overcoming the limitations of manual interpretation and enhancing the utility of wide-field imaging for tracking FECD progression.