Related Experiment Videos
Automated Identification of Diabetic Retinopathy Using Deep Learning
Rishab Gargeya1, Theodore Leng2
1The Harker School, San Jose, California.
Ophthalmology
|April 1, 2017
Summary
An artificial intelligence tool can automatically screen diabetic retinopathy (DR) from fundus images, achieving high accuracy. This technology aids in early detection and referral, potentially preventing vision loss in diabetic patients.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide.
- Many diabetic patients lack timely retinal screening, leading to undiagnosed and untreated cases.
- Automated screening can address the unmet need for widespread DR detection.
Purpose of the Study:
- To develop a robust diagnostic technology for automated diabetic retinopathy screening.
- To create an AI algorithm capable of classifying fundus images for DR detection.
- To facilitate timely referral of patients with DR to ophthalmologists.
Main Methods:
- Developed and evaluated a data-driven deep learning algorithm for DR detection.
- Trained and tested the AI model on 75,137 public fundus images.
- Validated the model using the MESSIDOR 2 and E-Ophtha databases.
Main Results:
- The AI model achieved a 0.97 AUC with 94% sensitivity and 98% specificity on local data.
- External validation on MESSIDOR 2 and E-Ophtha databases yielded AUC scores of 0.94 and 0.95.
- The algorithm generated abnormality heatmaps to highlight regions of interest for clinical review.
Conclusions:
- A data-driven AI grading algorithm can reliably screen fundus photographs for DR.
- The AI tool effectively identifies diabetic patients requiring ophthalmologist referral.
- Global implementation of this AI algorithm could significantly reduce vision loss from DR.