For some time now, AI programs have been trained to detect diseases from images of the eye fundus. What's unique about the new tool (RETFound), developed by Pearse Keane's lab in London, is that not every one of the 1.6 million retinal images used for training had to be labeled as 'normal' or 'abnormal' by a human. Such processes are time-consuming but often required in the development of standard machine learning models.
In their work published in Nature, the authors used a method also employed for training large language models like ChatGPT. For ChatGPT, countless examples of human-generated texts are used to learn how to predict the next word in a sentence based on the context of the preceding words. Similarly, RETFound uses retinal images to learn what missing sections of the retina might look like. After reviewing millions of images, the model then learns what the retina should look like at each location.
The retina, easily observable with an ophthalmoscope, thus provides a window into our health since it is the only part of the human body where a capillary network consisting of the smallest blood vessels can be directly observed. This is particularly useful in diagnosing conditions like hypertension or diabetes. Moreover, the retina can also be considered an extension of the central nervous system (i.e. the diencephalon), indicating similarities with the brain, which means that retinal images can also be used for diagnosing neurological diseases.
After Keane and colleagues had pretrained RETFound on about 1.6 million unlabeled retinal images, 100 images from individuals who had developed Parkinson's and 100 from control persons were entered into the software, which could then quickly learn the retinal features possibly associated with Parkinson's. Thus, only 100 human interactions were needed to potentially make a correct Parkinson's diagnosis on retinal images read by the machine.
The system has proven very effective in detecting eye diseases such as diabetic retinopathy. On a scale where 0.5 represents a model no better than random prediction, and 1 represents a perfect model making accurate predictions every time, it scores between 0.82 and 0.94. While the overall performance in predicting the risk for systemic diseases like heart attacks, strokes, or Parkinson's is still limited, it is clearly better than previous AI models. Particularly interesting in this context is the application of RETFound to magnetic resonance imaging or CT scans.
It seems certain that AI will play a central role in the early diagnosis of neurodegenerative diseases in the future.
References:
Zhou Y, Chia MA, Wagner SK, …, Keane PA (2023) A foundation model for generalizable disease detection from retinal images. Nature 622:156
Lenharo M (2023) AI detects eye disease and risk of Parkinson’s from retinal images. Nature News, doi.org/10.1038/d41586-023-02881-2
Image credit: iStock/James Ebanks
Commentaires