International Business Department           Liu Bojia           September 15, 2023

  We often say that our eyes are our windows to the world. What is less well known is that at the same time, our eyes can also be the window to the outside world for insight into our health.

  It is the retina, located in the back wall of the eye, that records the state of our health. The retina is unique in that it is the only structure in the body where the capillary network can be directly observed. When we suffer from systemic cardiovascular disease (e.g. hypertension), every blood vessel in the body may be affected, and pathologically relevant changes can therefore be reflected in the capillary network of the retina. In addition, as an extension of the central nervous system, the retina contains information about the neural tissue, providing the possibility of non-invasive visualization of the nervous system.

  Understanding how the eye is connected to the body has important implications for coping with complex diseases as well as problems associated with aging. But to the naked eye, these intricate networks are a jumbled mess from which we simply cannot read valid health information. Luckily, advances in artificial intelligence (AI) offer opportunities.

  In a recent study published in Nature, scientists from University College London and Moorfields Eye Hospital have developed an AI retina-based model called RETFound, which can diagnose sight-threatening eye diseases based on images of people's retinas and is expected to be used to predict the risk of a wide range of systemic diseases such as cardiovascular disease and Parkinson's disease. Yukun Zhou of University College London and Moorfields Eye Hospital is the first author and co-corresponding author of the paper.

  Prior to this study, AI tools that detect diseases through retinal images have been available. But the biggest problem with such traditional AI models is that they often require a large number of images with high-quality labels for training, and the addition of labels requires a huge amount of specialized physician workload, as well as high costs.

  In contrast, the latest RETFound tool was developed using a self-supervised learning strategy. Using an approach similar to training a large language model such as ChatGPT, the team used a large number of unlabeled retinal images to learn how to predict the missing parts of an image. The two base models of the latest research, the RETFound model, were used for color fundus photography as well as optical coherence tomography. Such base models are characterized by the ability to be fine-tuned for a range of subsequent detection and prediction tasks after being trained on a large amount of unlabeled data, with the potential to be adapted to an infinite number of application scenarios.

  RETFound was able to match the performance of other AI systems using only as low as 10% of human labels, achieving a huge increase in efficiency. Subsequent tests have also validated RETFound's good results in the diagnosis of eye diseases such as diabetic retinopathy and glaucoma. Among other things, in detecting diabetic retinopathy, the scores ranged between 0.822 and 0.943, depending on the dataset used (a score of 0.5 indicates no better than randomized prediction; a score of 1 indicates a perfect model that makes accurate judgments every time).

  In addition, the study demonstrated the potential for using RETFound to predict more systemic diseases. The authors first pretrained RETFound with 1.6 million unlabeled retinal images. Then, using only a small number of images, such as retinal images of 100 people with Parkinson's disease versus 100 people without the disease, the model was able to learn features of the retina that are associated with Parkinson's disease. Based on this strategy, RETFound has demonstrated superior performance to other AI models in predicting systemic diseases such as heart failure, myocardial infarction, stroke and Parkinson's disease.

  Prof. Pearse Keane, who led the study, said that the study demonstrates several scenarios where RETFound can be used, but it has the potential to be further developed and applied to hundreds of eye diseases that threaten vision that we have not yet explored.

  The model is currently open source (https://github.com/rmaphoh/RETFound_MAE). The team hopes that more researchers will be able to apply RETFound to a wider range of application scenarios by training and fine-tuning it. In the meantime, the research team is looking ahead to see if the techniques used to develop RETFound can be generalized to more complex medical images, such as magnetic resonance images or computed tomography scans, to provide tools for diagnosing or predicting a wider range of diseases.