Severity Classification of Conjunctival Hyperaemia by Deep Neural Network Ensembles

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Conjunctival hyperaemia is a common clinical sign in ophthalmology.
• Existing severity classifications lack objective criteria.
• Objective grading of conjunctival hyperaemia is needed for accurate diagnosis and management.

Purpose of the Study:

• To develop and validate a deep learning-based software for objective conjunctival hyperaemia grading.
• To assess the performance of neural networks in classifying conjunctival hyperaemia severity.
• To establish a consistent and rapid method for evaluating ocular redness.

Main Methods:

• Utilized 3700 images for training and 923 for validation of nine neural network models.
• Employed deep learning architectures including DenseNet201, DenseNet121, VGG19, and ResNet50.
• Evaluated model performance and selected the best performing combination for the grading system.

Main Results:

• The DenseNet201 model demonstrated superior individual performance.
• A combination of DenseNet201, DenseNet121, VGG19, and ResNet50 achieved the best results.
• A high correlation (0.737, p < 0.01) was found between the system's output and vessel-area occupation.

Conclusions:

• The developed AI system provides objective and consistent grading of conjunctival hyperaemia.
• This deep learning approach offers comparable accuracy to specialists but with increased speed.
• The software represents a significant advancement in the objective assessment of ocular disorders.