Related Experiment Video
Updated: Oct 25, 2025
04:45
An Epithelial Abrasion Model for Studying Corneal Wound Healing
Published on: December 29, 2021
3.5K
Differentiation of Active Corneal Infections from Healed Scars Using Deep Learning
Mo Tiwari1, Chris Piech1, Medina Baitemirova2
1Department of Computer Science, Stanford University, Stanford, California.
Ophthalmology
|August 5, 2021
Summary
An automated algorithm using a convolutional neural network (CNN) accurately differentiates active corneal ulcers from scars using photographs. This AI tool shows promise for accessible eye care diagnostics in underserved regions.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Corneal ulcers are a significant cause of vision impairment globally.
- Accurate differentiation between active corneal ulcers and healed scars is crucial for appropriate treatment and visual outcome.
- Current diagnostic methods may require specialized equipment and expertise, limiting accessibility in certain settings.
Purpose of the Study:
- To develop and validate an automated, portable algorithm for distinguishing active corneal ulcers from healed scars using only external eye photographs.
- To assess the algorithm's accuracy and generalizability across diverse patient populations.
Main Methods:
- A convolutional neural network (CNN) was trained on a dataset of 1313 corneal ulcers and 1132 corneal scars from multiple clinical trials and institutions.
- The trained CNN was tested on independent datasets from eye clinics in India (n=200) and Stanford University (n=101).
- Performance was evaluated using F1 score, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve, with feature importance visualized using gradient-weighted class activation mapping.
Main Results:
- The CNN achieved high accuracy in classifying active corneal ulcers and scars in both testing datasets.
- In the Indian cohort, the CNN demonstrated an F1 score of 92.0% and an AUC of 0.9731.
- In the Stanford cohort, the CNN achieved an F1 score of 84.3% and an AUC of 0.9474, with visualizations correlating to clinical features like infiltrate and hypopyon.
Conclusions:
- The developed CNN algorithm accurately differentiates active corneal ulcers from scars and generalizes well to new patient populations.
- The algorithm's ability to focus on clinically relevant features suggests its utility in diagnosis.
- This automated approach holds potential as an inexpensive and accessible tool for aiding triage in communities with limited access to eye care.
