Updates in Diagnostic Imaging for Infectious Keratitis: A Review

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Infectious keratitis (IK) is a leading cause of global blindness, necessitating prompt diagnosis and treatment to prevent vision loss.
• Current diagnostic methods include slit lamp microscopy, corneal scraping cultures, and advanced imaging techniques like optical coherence tomography (OCT) and in vivo confocal microscopy (IVCM).
• IVCM demonstrates high sensitivity and specificity for diagnosing fungal and Acanthamoeba keratitis, while OCT aids in assessing infection severity and progression.

Purpose of the Study:

• To explore the potential of deep learning (DL) models as diagnostic aids for infectious keratitis (IK) using image recognition.
• To review the current state of DL applications in IK diagnosis, including methodologies and limitations.
• To identify future directions for improving DL models for IK diagnosis, particularly in real-world clinical settings.

Main Methods:

• Review of studies developing DL models for IK diagnosis, primarily utilizing slit lamp photographs.
• Analysis of different DL approaches, including single convolutional neural network (CNN) algorithms and ensemble methods.
• Examination of imaging techniques such as OCT and IVCM as complementary diagnostic tools.

Main Results:

• DL models, particularly those using CNNs trained on slit lamp images, show promise in aiding IK diagnosis.
• IVCM offers high diagnostic accuracy for specific types of keratitis, such as fungal and Acanthamoeba infections.
• Limitations of current DL models include the need for large datasets, feature extraction challenges, data imbalance, and misclassification bias.

Conclusions:

• Deep learning holds significant potential for improving the diagnosis of infectious keratitis, especially when leveraging slit lamp photography.
• Further research and development are essential to address the limitations of DL models, including data requirements and algorithmic biases.
• Emerging AI technologies like generative adversarial networks may offer solutions for enhancing DL model performance and applicability in clinical practice.