Related Experiment Video
Updated: Aug 6, 2026
06:53
A Custom Multiphoton Microscopy Platform for Live Imaging of Mouse Cornea and Conjunctiva
Published on: May 17, 2020
Triple-Phase Multimodal Knowledge Aggregation Framework for Microbial Keratitis Subtype Diagnosis on Slit-Lamp
Yiqing Wang1, Maria A Woodward2, Ziyun Yang1
1Department of Biomedical Engineering, Duke University, Durham, NC, USA.
Arxiv
|July 17, 2026
Summary
A new AI framework accurately distinguishes bacterial from fungal keratitis using eye photos and clinical data. This rapid diagnostic tool aids timely treatment for microbial keratitis, outperforming existing methods.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Microbial keratitis diagnosis relies on slow, resource-intensive culture and PCR methods.
- Rapid pathogen identification is crucial for effective microbial keratitis treatment.
Purpose of the Study:
- To develop and evaluate a novel AI framework for rapid bacterial versus fungal keratitis classification.
- To improve diagnostic speed and accuracy for microbial keratitis.
Main Methods:
- A triple-phase multimodal AI framework was developed using slit-lamp photographs (blue-light, sclerotic-scatter, white-light) and clinical metadata.
- The model employed cross-modality contrastive learning, modality-specific fine-tuning, and multimodal ensemble learning.
- Evaluation was conducted on a multicenter dataset of 1,645 patients from India and the US.
Main Results:
- The framework achieved 85.84% accuracy, 84.46% average F1-score, and 0.885 AUC.
- Site-specific evaluation highlighted the need for realistic cross-site generalization assessment.
- The developed framework demonstrated superior performance compared to other approaches.
Conclusions:
- The AI framework offers a rapid and accurate method for differentiating bacterial from fungal keratitis.
- This approach has the potential to significantly improve clinical decision-making and patient outcomes in microbial keratitis management.
- Further validation and implementation could overcome current diagnostic limitations.
