Development of a Deep Learning Algorithm for Automatic Diagnosis of Diabetic Retinopathy

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy is a leading cause of blindness, necessitating early detection through funduscopic screening.
• Current diagnostic methods face delays, financial costs, and risks of vision loss.
• Automated analysis of funduscopic images offers a potential solution to these challenges.

Purpose of the Study:

• To apply deep learning, specifically convolutional neural networks, for automated diabetic retinopathy staging.
• To develop a deep learning model for detecting eye laterality (left or right) from funduscopic images.
• To evaluate the performance of these models on large datasets.

Main Methods:

• A convolutional neural network (CNN) architecture was employed for image classification tasks.
• The models were trained on a large, publicly available Kaggle dataset of funduscopic images.
• Performance was assessed using sensitivity, specificity, and accuracy metrics on separate validation datasets.

Main Results:

• The diabetic retinopathy staging model achieved 80.28% sensitivity and 92.29% specificity on a validation set of approximately 53,000 images.
• The eye laterality detection model achieved 93.28% accuracy on a validation set of 8,816 images.
• Training involved approximately 35,000 images for retinopathy staging and 8,810 images for laterality detection.

Conclusions:

• Deep learning models demonstrate significant potential for accurate and efficient automated diagnosis of diabetic retinopathy.
• The developed models can effectively classify disease stage and determine eye laterality, aiding ophthalmologists.
• This technology could improve early detection rates, reduce diagnostic delays, and mitigate the risk of blindness in diabetic patients.