Related Experiment Video
Updated: May 5, 2026
09:22
Establishing a Porcine Ex Vivo Cornea Model for Studying Drug Treatments against Bacterial Keratitis
Published on: May 12, 2020
5.2K
Multimodal Deep Learning for Differentiating Bacterial and Fungal Keratitis Using Prospective Representative Data.
N V Prajna1, Jad Assaf2, Nisha R Acharya3
1Aravind Eye Hospital, Madurai, India.
Ophthalmology Science
|January 27, 2025
Summary
A computer vision model trained on South Indian keratitis data achieved high accuracy in differentiating bacterial and fungal infections. This highlights the potential of AI in diagnosing infectious keratitis using representative datasets.
Area of Science:
- Ophthalmology
- Medical Artificial Intelligence
- Infectious Diseases
Background:
- Infectious keratitis, a major cause of vision loss, requires accurate differentiation between bacterial and fungal etiologies for effective treatment.
- Current diagnostic methods can be time-consuming and may not always be definitive, necessitating improved diagnostic tools.
Purpose of the Study:
- To develop and evaluate multimodal machine learning models for differentiating bacterial and fungal keratitis.
- To compare the performance of models trained on a prospective, representative dataset versus a dataset from prior clinical trials.
Main Methods:
- Development of three models: clinical data, computer vision (EfficientNet), and multimodal.
- Training and validation using a prospective dataset (MADURAI) from South India (599 subjects).
- Comparison of a model trained on the MADURAI dataset against one trained on prior randomized clinical trial (RCT) data.
Main Results:
- The computer vision model trained on the MADURAI dataset demonstrated superior performance (AUPRC 0.94) compared to the clinical data model and the RCT-trained computer vision model.
- The multimodal model did not offer significant performance improvement over the computer vision model.
- Key performance metrics included AUPRC, AUROC, accuracy (77%), and F1 score (0.85).
Conclusions:
- A computer vision model trained on a prospective, representative dataset is highly effective for diagnosing infectious keratitis.
- Image-based deep learning shows promise for enhancing diagnostic capabilities in infectious keratitis.
- The study underscores the critical importance of using representative, real-world data for training and evaluating machine learning models in healthcare.
