Opacity Detection on Optical Coherence Tomography Based on an Incidence-Angle and Depth-Dependent Model of Corneal Reflectance.

PURPOSE: To develop and validate an automated corneal opacity detection algorithm for optical coherence tomography (OCT) images, utilizing an incidence-angle- and depth-dependent model of corneal reflectance.

DESIGN

Retrospective, cross-sectional diagnostic accuracy study.

SUBJECTS: Training used 95 healthy eyes from 49 volunteers. Testing included 50 eyes from 42 patients with corneal opacities and 35 healthy eyes from 35 volunteers.

METHODS

Normal-eye OCT scans were used to model normative incidence-angle-dependent reflectance across corneal layers. The algorithm detected pixels above the normal reflectance range using model-based thresholds, binned percentile analysis, and morphological operations. Eye-level performance was evaluated against slit-lamp examination as clinical ground truth and compared with 5 trained physician annotators. Pixel-level agreement with consensus annotations (≥3 of 5 annotators) was assessed with Dice similarity coefficient.

MAIN OUTCOME MEASURES: Eye-level accuracy, F1-score, sensitivity, and specificity; pixel-level Dice similarity coefficient and segmented-area agreement versus consensus annotations.

RESULTS

At the eye level, the algorithm achieved accuracy of 0.93, F1-score of 0.94, sensitivity of 0.96, and specificity of 0.89. Human annotators had a mean accuracy of 0.83 ± 0.06, F1-score of 0.85 ± 0.04, sensitivity of 0.84 ± 0.09, and specificity of 0.80 ± 0.27. At the pixel level, mean Dice similarity coefficient versus consensus was 0.58 for the algorithm and 0.71 ± 0.05 for annotators. The algorithm's total segmented opacity area was close to the consensus pixel count (98% of consensus).

CONCLUSION

An algorithm that incorporates incidence angle and depth-specific reflectance thresholds detects and segments corneal opacities. It demonstrated favorable accuracy at the eye level and produced quantitative opacity maps on OCT.