Deep Learning Analysis of Widefield Cornea Endothelial Imaging in Fuchs Dystrophy

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

Cross-sectional clinical observational study.

A total of 1839 images were obtained via WFSM imaging (CEM-530, Nidek Co Ltd) performed on 155 FECD eyes. A separate data set comprising images from 50 FECD eyes and 50 control eyes (70% training, 30% validation) was used for DLN training, which was tested on 354 images from 55 eyes from varying regions (central, paracentral, and peripheral).

Images were graded based on a standardized quality score. Central images were compared with paracentral and peripheral images in terms of quality and morphometric parameters: endothelial cell density (ECD), coefficient of variation (CV), and hexagonality (HEX). A U-Net-based DLN was developed and trained using the separate data set and then tested on an external longitudinal data set (baseline and 1 month). Segmentation accuracy between DLN and manual analysis was compared using the Sørensen-Dice coefficient. Morphometric outcomes (ECD, HEX, and CV) were analyzed using paired t tests.

Intergrader agreement for image quality and FECD severity; comparison of DLN-derived ECD with manual analysis.

Strong intergrader agreement was observed for both image quality (κ = 0.967, 95% confidence interval [CI]: 0.959-0.976) and FECD severity (κ = 0.891, 95% CI: 0.786-0.995). Endothelial cell density differences between paracentral/peripheral regions were significant in eyes without or with subclinical edema (P = 0.001-0.011). Deep learning network-based segmentation closely matched manual results (Dice coefficient = 0.86 ± 0.04). Central ECD values obtained via DLN were significantly higher than manual analysis (DLN: 2633.12 ± 1167.3 cells/mm² vs. manual: 1728.58 ± 805.69 cells/mm², P < 0.001).

This study presents a novel application of deep learning for analyzing widefield corneal endothelial images. The integration of a progression visualization tool enhances interpretability, allowing efficient autoanalysis and organization of large WFSM data sets-streamlining workflows and addressing limitations of manual interpretation.

The authors have no proprietary or commercial interest in any materials discussed in this article.