Introducing RETFound, an AI that detects signs of eye disease and Parkinson's disease from retinal images, and self-supervised learning simplifies pre-training

AI that detects diseases using retinal images already exists, but it required training by first labeling a large number of retinal images as ``normal'' or ``not normal.'' The new AI ' RETFound ' performs 'self-supervised learning' similar to the mechanism of ChatGPT, so training in advance will be quite easy.

A foundation model for generalizable disease detection from retinal images | Nature

AI detects eye disease and risk of Parkinson's from retinal images

The retina is the only part of the human body where the capillary network can be directly observed, so examining the retina can tell you about a person's health status.

Scientists are developing a technology using AI to detect signs of disease from retinal images, but this requires labeling retinal images as ``normal'' or ``abnormal.'' Additional training was required.

The newly developed 'RETFound' eliminates the need for labeling by performing self-supervised learning. Self-supervised learning is something that is used in chat AI, such as ChatGPT. In the case of ChatGPT, it predicts the next word from the context of the previous word, using a vast amount of human-generated text as precedent.

Ophthalmologist Pierce Keene, co-author of the paper, trained RETFound on 1.6 million unlabeled retinal images. Through this, RETFound learns 'what the retina should look like'.

Then, by inputting a small number of labeled images, such as ``100 retinal images of people with Parkinson's disease and 100 retinal images of people without Parkinson's disease'', RETFound can detect signs of disease that appear on the retina. He said he was able to learn easily.

RETFound has already shown excellent results in detecting conditions such as diabetic retinitis. Although its performance is lower in predicting systemic diseases such as heart attack, heart failure, stroke, and Parkinson's disease than in diabetic retinitis, it seems to be better than other AIs.

Liu Xiaoxuan, a clinical researcher at the University of Birmingham, commented that RETFound 'removes a major bottleneck for researchers' by first using and training on unlabeled data. 'So far, this is one of the few successful applications of the basic model to medical images,' he said.