A Convolutional Neural Network Using Anterior Segment Photos for Infectious Keratitis Identification

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Corneal infections, including bacterial and fungal keratitis, pose significant threats to vision.
• Accurate and timely diagnosis is crucial for effective treatment and prevention of vision loss.
• Distinguishing between infectious and non-infectious corneal conditions can be challenging.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for differentiating bacterial keratitis, fungal keratitis, non-infectious corneal lesions, and normal corneas.
• To assess the performance of convolutional neural networks (CNNs) and ensemble methods in corneal image analysis.

Main Methods:

• A retrospective study utilizing 6,478 slit-lamp photos from 2,171 eyes.
• Development of deep learning models including ResNet50, DenseNet121, and VGG19.
• Application of an ensemble technique with probability weighting for improved diagnostic accuracy.

Main Results:

• DenseNet121 achieved an accuracy of 0.8 (95% CI 0.74-0.86) among individual CNNs.
• The ensemble technique demonstrated superior performance with an accuracy of 0.83 (95% CI 0.78-0.88).
• The algorithm showed high performance in discriminating between various corneal conditions.

Conclusions:

• Deep learning models, particularly ensemble techniques, show promise for accurate corneal condition diagnosis.
• The developed algorithm can serve as a valuable screening tool for healthcare providers.
• Facilitates rapid provisional diagnosis of infectious keratitis, aiding clinical decision-making.