A deep learning model for identifying diabetic retinopathy using optical coherence tomography angiography.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetes prevalence is rising, increasing the need for widespread diabetic retinopathy (DR) screening.
• Traditional DR screening methods are labor-intensive and time-consuming.
• Advances in AI offer potential for automated, accurate DR detection from retinal images.

Purpose of the Study:

• To develop and validate a fully automated classification algorithm for diagnosing DR and identifying referable status.
• To assess the performance of a convolutional neural network (CNN) model using optical coherence tomography angiography (OCTA) images for DR classification.
• To compare the CNN model's performance against conventional machine learning models.

Main Methods:

• Developed a fully automated DR classification algorithm using a CNN model on OCTA images.
• Established ground truths for classification using ultra-widefield fluorescein angiography for enhanced data annotation accuracy.
• Validated the CNN classifier's performance against conventional machine learning models and through external validation.

Main Results:

• The proposed CNN classifier achieved high accuracy (91-98%), sensitivity (86-97%), specificity (94-99%), and AUC (0.919-0.976).
• Similar performance was observed in external validation.
• Classification accuracy remained consistent across varying OCTA image sizes and depths, including narrow macular regions and single image slabs.

Conclusions:

• CNN-based classification of DR using OCTA images is feasible and highly accurate.
• The algorithm demonstrates potential for efficient DR screening and referral.
• This technology is expected to establish a novel diagnostic workflow for DR detection.