Machine Learning Using Anterior Segment Oct-Derived Indices For The Classification Of Fuchs Endothelial Corneal Dystrophy Severity

Fuchs endothelial corneal dystrophy (FECD) is a common, age-related, progressive corneal disease characterised by corneal guttae, endothelial decompensation with stromal oedema, and visual loss. Anterior segment optical coherence tomography (AS-OCT) can accurately measure parameters of corneal shape, such as pachymetry and surface elevation, with detailed cross-sectional images. Although numerous studies have demonstrated the utility of machine learning (ML) in ophthalmology, few have applied this to FECD. This study explores the potential of ML models trained on AS-OCT-derived indices to classify FECD disease severity. 

Moorfields Eye Hospital, London, United Kingdom.  

This observational case-control study included consecutive pre-operative FECD cases from January 2020 to November 2023. Cases were categorized clinically as early-stage (only guttae) or late-stage (with corneal oedema). We used the healthy corneas of patients attending laser-refractive surgery as the control group. We also included eyes with pseudophakic bullous keratopathy (PBK) for comparison with late-stage FECD. Indices derived from AS-OCT scans (MS39, CSO, Italy) were used to develop ML classification models. The data were split 75% for training and 25% for validation. We calculated the area under the curve (AUC), specificity, and sensitivity of the prediction outputs. 

We included 1755 eyes (976 controls, 108 early-stage FECD, 230 late-stage FECD, and 441 PBK) to develop the classification models. In the validation set, the models achieved an AUC of 0.96 with 76% sensitivity and 99% specificity in classifying early-FECD from controls; an AUC of 0.87 with 96% sensitivity and 63% specificity in classifying late-FECD and early-FECD; and AUC of 0.87 with 92% sensitivity and 71% specificity in classifying late-FECD and PBK. The model highlighted that deviation of the location of the minimal corneal thickness was the most sensitive parameter for distinguishing early-FECD from healthy cornea, whilst central corneal thickness was key in distinguishing early- and late-FECD.  

This study demonstrates that ML trained on AS-OCT-derived indices can achieve comparable diagnostic performance to published studies trained on raw images, despite only requiring a fraction of computational resources. This proof-of-concept application of ML underscores the potential of training models using AS-OCT-derived indices for corneal pathology diagnosis over raw images, therefore reducing reliance on resource-intensive datasets and holding promise for broader applications beyond FECD.