Diabetes detection from non-diabetic retinopathy fundus images using deep learning methodology

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetes is a major global health concern, leading to significant morbidity and mortality.
• Current diabetes detection from retinal images relies on identifying retinopathy, a late-stage complication.
• There is a need for non-invasive, early detection methods for diabetes.

Purpose of the Study:

• To develop and evaluate an AI machine learning model for detecting diabetes using fundus imagery.
• To assess the model's ability to detect diabetes independent of diabetic eye disease indicators.
• To explore the potential of fundus imagery for non-invasive diabetes diagnosis.

Main Methods:

• A machine learning algorithm was trained on the large-scale EyePACS dataset (47,076 images).
• The dataset included patient cohorts stratified by disease duration and healthy controls.
• Performance was evaluated using the area under the receiver operating curve (AUC).

Main Results:

• The AI model achieved an AUC of 0.86 for detecting diabetes per patient visit.
• The model demonstrated an AUC of 0.83 for detecting diabetes per image.
• Results indicate successful diabetes detection solely from fundus images, irrespective of retinopathy.

Conclusions:

• Fundus imagery holds potential for non-invasive diabetes diagnosis.
• This AI-driven approach could facilitate point-of-care and accessible diabetes screening.
• The technology may help diagnose previously undiagnosed individuals with diabetes.