Deep Learning Highly Accurate in Classifying Diabetic Retinopathy on OCT-A

Images obtained from optical coherence tomography angiography (OCT-A) can be evaluated by a deep convolutional neural network, according to research presented at the American Academy of Ophthalmology 2020 meeting, held virtually November 13 to 15. Deep convolutional neural networks are a deep learning algorithm that can accurately identify and classify images, in this case, of diabetic retinopathy and its severity. 

Researchers tested this technology using 2489 OCT-A images from 1163 eyes (857 eyes with diabetes and 306 controls). A deep convolutional neural network was trained to classify diabetic retinopathy stages using superficial, deep, and whole retinal layer images. Data were divided into 4 subsets, on which investigators performed 4-fold cross-validation. Neural network outputs were compared with the outputs from both retinal experts and traditional machine learning methods. 

Within the convolutional neural network, investigators identified an accuracy of 85% for 6-level staging tasks, compared with 53% accuracy of traditional machine learning methods and 31% accuracy of retinal experts. 

“The deep learning-based classification using OCT-A images can provide clinicians with a classification of [diabetic retinopathy] severity for patients with diabetes,” the researchers concluded. 

Reference

Sagong M, Cha S. A deep convolutional neural network for classification of DR using OCT-A. Presented at: American Academy of Ophthalmology 2020 Annual Meeting; November 13-15, 2020. Abstract PO400.