Artificial Intelligence Can Identify Anomalies From Diabetic Retinopathy

Retinal scan testing for glaucoma. Woman with her head resting in a machine (left) being used by an ophthalmologist (right) to scan the retinas of her eyes and examine them for signs of glaucoma. The retina is the light-sensitive layer at the back of the eye responsible for vision. Glaucoma is a build-up of pressure inside the eye causing blurring and blindness. The technique in use here is optical coherence tomography (OCT) using a confocal scanning laser ophthalmoscope (SLO) device. This machine is from the Optovue company. The results of the scan are shown on the screens and in image C028/1548.
A study suggests anomaly detectors not trained with diseased retina can detect diabetic retinopathy.

Can artificial intelligence (AI) systems trained on normal data recognize anomalies in retinal images? According to research published in JAMA Ophthalmology, AI-based anomaly detectors can be used to screen for retinal diseases. 

Advancements in anomaly detection algorithms using machine learning and AI techniques have only recently been applied to the field of ophthalmology, according to researchers. For example, anomaly detection has been used in optical coherence tomography (OCT) segmentation, fundus image drusen delineation, brain magnetic resonance imaging segmentation, myopathy detection by ultrasonography, and fundus screening. However, unavailable annotations or lack of training examples for certain diseases, such as rare diseases, novel phenotypes, or rare variants of existing diseases, are challenging for deep learning systems. To clear this hurdle, anomaly detection only requires training with normal images and can flag anomalies or previously unknown presentations of diseases.

This cross-sectional study compared 16 variations of anomaly detectors using a surrogate problem including non-referable and referrable diabetic retinopathy. Only retinas with nonreferable diabetic retinopathy (no diabetic macular edema, no diabetic retinopathy, mild to moderate nonproliferative diabetic retinopathy) were used to train the AI system. Then, both nonreferable and referable diabetic retinopathy (including diabetic macular edema or proliferative diabetic retinopathy) were used to test how well the system detected retinal disease.

The study used 88,692 high-resolution retinal images of 44,346 patients with diabetic retinopathy in various stages from the EyePACS data set. The best performing across all anomaly detectors had an area under the receiver operating characteristic of 0.808 (95%CI, 0.789-0.827) and was obtained using an embedding method that involved a self-supervised network, according to investigators.

“Anomaly detectors could be pursued for retinal diagnoses based on artificial intelligence systems that may not have access to training examples for all retinal diseases in all phenotypic presentations,” according to the investigators. 

Possible applications include screening the population for any retinal disease rather than a specific disease, detecting novel retinal diseases or novel presentations of common retinal diseases, and detecting rare diseases with little or no data available for training.

Limitations of this study, according to its authors, included using a surrogate problem examining nonreferrable vs referrable diabetic retinopathy only and EyePACS clinical data only.


Burlina P, Paul W, Liu TYA, et al. Detecting anomalies in retinal diseases using generative, discriminative, and self-supervised deep learning. JAMA Ophthalmol. Published online December 30, 2021. doi:10.1001/jamaophthalmol.2021.5557