Ensemble Machine Learning Approaches for Automated Fungal Keratitis Diagnosis Using In Vivo Confocal Microscopy...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Fungal keratitis (FK) is a severe eye infection risking vision loss.
• Prompt diagnosis and treatment are vital for managing FK.
• Current diagnostic methods can be slow and resource-intensive.

Purpose of the Study:

• To evaluate state-of-the-art machine learning techniques for classifying FK using in vivo confocal microscopy (IVCM) images.
• To assess the impact of various image processing and model tuning strategies on FK detection accuracy.
• To identify robust models adaptable to clinical settings and imbalanced datasets.

Main Methods:

• Systematic evaluation of machine learning models for FK classification from IVCM images.
• Experimentation with diverse image processing techniques, data augmentation, and hyperparameter tuning.
• Performance assessment focusing on accuracy, F1-scores, and model adaptability.

Main Results:

• A Random Forest model with green channel preprocessing and a 12-feature set achieved 99% accuracy in FK detection.
• Complex methods like histogram modeling yielded lower accuracy (64%).
• AdaBoost and RUSBoost models demonstrated robustness and high F1-scores, suitable for imbalanced datasets.

Conclusions:

• Machine learning analysis of IVCM images offers a highly accurate method for fungal keratitis detection.
• Specific preprocessing techniques (green channel) and feature sets significantly enhance diagnostic performance.
• Robust models like AdaBoost and RUSBoost show promise for real-world clinical application in FK diagnosis.