Automated vs. human evaluation of corneal staining

Area of Science:

• Ophthalmology
• Medical Imaging Analysis
• Computational Biology

Background:

• Corneal fluorescein staining is a key diagnostic for dry eye disease (DED).
• Human grading of corneal staining lacks consistency, impacting diagnostic reliability.
• A need exists for automated, objective quantification of corneal staining.

Purpose of the Study:

• To develop a software-assisted grading algorithm for corneal staining.
• To compare the algorithm's performance against human graders with varying experience levels.
• To assess the consistency and reliability of automated versus manual DED diagnosis.

Main Methods:

• Developed an ImageJ-based algorithm to detect and count superficial punctate keratitis.
• Trained the algorithm on 20 standardized corneal images from DED patients.
• Validated the algorithm on a separate dataset of 30 images, compared with 22 ophthalmologists using the Oxford grading scheme.

Main Results:

• The algorithm showed significant correlation (Sr=0.91, p<0.01) with expert-assessed corneal staining severity.
• Human graders exhibited moderate intrarater agreement (Kappa=0.426), while the software achieved perfect consistency (Kappa=1.0).
• Interrater agreement among humans varied significantly with experience, from 25.6% to 75.6%.

Conclusions:

• While human grading is adequate for clinical practice, software-assisted grading provides superior consistency.
• Automated grading is preferable for objective patient re-evaluation, especially in clinical trials.
• The developed algorithm enhances the reliability of corneal staining assessment in DED.