Corneal Opacity Detection On Optical Coherence Tomography Based On Corneal Reflectivity Profile

To develop and validate a novel algorithm for automated corneal opacity detection and segmentation in optical coherence tomography (OCT) images, leveraging the corneal reflectivity profile.

This study was conducted at the Casey Eye Institute, Oregon Health & Science University, Portland, OR.

338 normal OCT scans (149 eyes; 76 patients) were used to analyze the normative corneal intensity profile, considering layer and incidence angle variations. Superpixels were employed to streamline the analysis, setting opacity detection thresholds at the 99^th intensity percentile for each superpixel for medium and high opacities.

We validated our algorithm on 26 normal control OCT scans and 54 OCT scans (27 eyes; 14 patients) with different corneal opacities, including corneal scars, granular dystrophy, Reis-Bucklers corneal dystrophy, Salzmann nodular degeneration, keratoconus, and others. We compared the resultant opacities against manual segmentations by three experts, using the Dice score, which compares the intersection over the union.

Overall dice score for all OCT scans between professional annotators was 65%, which suggests poor agreement between the annotators manual labeling. The algorithm scored a dice score of 55% with the annotators. Additional comparative metrics will be explored in the manuscript.

We developed an algorithm capable of objective detection and segmentation of corneal opacities based on the normative corneal reflectivity profile. This algorithm is not subject to operator dependent variability and can enhance the clinician’s assessment of corneal disease severity and prognosis.