Developing a Deep Learning Model to Evaluate Bulbar Conjunctival Injection with Color Anterior Segment Photographs

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Bulbar conjunctival injection is a key indicator of ocular surface disease.
• Accurate grading of conjunctival injection is crucial for diagnosis and treatment monitoring.
• Current grading methods can be subjective and time-consuming.

Purpose of the Study:

• To assess the feasibility of a deep learning (DL) model for grading bulbar conjunctival injection.
• To develop and validate a DL model using convolutional neural networks (CNNs).
• To compare DL model performance against human expert grading.

Main Methods:

• Collected 1401 anterior segment photographs.
• Utilized two CNN-based DL models for grading.
• Established ground truth using ophthalmologist-labeled injection scores.
• Evaluated model performance using accuracy, precision, recall, F1-score, Kappa, and AUC.

Main Results:

• The DL model achieved high performance metrics.
• Micro-average and macro-average AUC values were both 0.98.
• Accuracy reached 87.12%, precision 87.13%, recall 87.12%, and F1-score 87.07%.
• Cohen's Kappa was calculated at 0.8153, indicating strong agreement.

Conclusions:

• The DL model demonstrates excellent performance in grading bulbar conjunctival injection severity.
• This AI approach has potential for objective evaluation of ocular surface conditions.
• The model may aid in assessing disease progression and recovery in patients.