Corneal Epithelial Thickness Mapping in the Diagnosis of Ocular Surface Disorders Involving the Corneal Epithelium: A Comparative Study.

PURPOSE: The purpose of this study was to analyze the role of corneal epithelial thickness (ET) mapping provided by spectral domain optical coherence tomography in the diagnosis of ocular surface disorders (OSDs) involving the corneal epithelium. DESIGN: This was a retrospective comparative study. METHODS: Institutional settings are as follows. Study population includes 303 eyes with an OSD and 55 normal eyes (controls). Observation procedures include spectral domain optical coherence tomography with epithelial mapping in the central 6 mm. Main outcome measures include ET map classification (normal, doughnut, spoke-wheel, localized/diffuse, and thinning/thickening patterns) and ET data and statistics (minimum, maximum, and SD). A quantitative threshold was determined with receiver operating curves to distinguish pathological from normal corneas. Sensitivity and specificity of classification and quantitative data were calculated using all eyes to assess the ability to distinguish corneas with a given corneal disorder from other conditions. RESULTS: Classification of full agreement between 3 readers was obtained in 75.4% to 99.4% of cases. Main OSD features were keratoconus (135 eyes), doughnut pattern (sensitivity/specificity = 56/94%), and max-min ET ≥ 13 µm (84/43%); limbal deficiency (56 eyes), spoke-wheel pattern (66/98%), and max-min ET ≥ 14 µm (91/59%); epithelial basement membrane dystrophy (55 eyes), inferior thickening pattern (55/92%), and central ET > 56 µm (53/81%); dry eye (21 eyes), superior thinning pattern (67/88%), and minimal ET ≤ 44 µm (86/48%); pterygium (10 eyes), nasal thickening pattern (100/86%), and nasal ET > 56 µm (80/71%); and in situ carcinoma (11 eyes), max ET > 60 µm (91/60%), and ET SD >5 µm (100/58%). CONCLUSIONS: The epithelial map pattern recognition combined with quantitative analysis of ET is relevant for the diagnosis of OSDs and for distinguishing various OSDs from each other. Deep learning analysis of big data could lead to the fully automated diagnosis of these disorders.