Deep Learning Analysis of Widefield Cornea Endothelial Imaging in Fuchs Dystrophy

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Fuchs endothelial corneal dystrophy (FECD) affects corneal endothelium.
• Accurate analysis of widefield specular microscopy (WFSM) images is crucial for FECD assessment.
• Manual analysis of WFSM images can be time-consuming and subjective.

Purpose of the Study:

• To evaluate the efficacy of a deep learning network (DLN) for analyzing WFSM images in eyes with FECD.
• To compare DLN performance against manual analysis for key morphometric parameters.

Main Methods:

• A U-Net-based DLN was developed and trained on a dataset of WFSM images from FECD and control eyes.
• The DLN was tested on an independent dataset, analyzing central, paracentral, and peripheral corneal regions.
• Segmentation accuracy was assessed using the Sørensen-Dice coefficient, and morphometric outcomes (ECD, CV, HEX) were compared to manual analysis.

Main Results:

• The DLN demonstrated strong agreement with manual analysis, achieving a Dice coefficient of 0.86.
• DLN-derived endothelial cell density (ECD) was significantly higher than manual measurements (2633 vs. 1729 cells/mm²).
• Significant differences in ECD were observed between corneal regions in eyes with and without subclinical edema.

Conclusions:

• Deep learning offers a novel and efficient approach for analyzing WFSM images in FECD.
• The DLN tool enhances interpretability and streamlines the workflow for large datasets.
• This AI application addresses limitations of manual interpretation, improving diagnostic efficiency.