Researchers have developed an exciting new AI tool to analyze retinal scans and detect various diseases, from eye conditions to Parkinson’s. The new model, RETFound, was created using a training method that removes time-consuming and expensive manual labeling of medical images.
RETFound capitalizes on the retina’s unique window into human health.
- The retina is an extension of the brain and central nervous system, allowing direct observation of blood vessels and neural tissue.
- Retinal changes can indicate developing issues like cardiovascular disease, diabetes, or neurodegenerative disorders.
In tests, RETFound detected vision-threatening diseases like diabetic retinopathy and the risk of Parkinson’s and stroke.
Self-supervised learning to streamline medical AI training could be game-changing.
High-quality labels for medical data are extremely expensive, so label efficiency has become the coin of the realm. —Curtis Langlotz, M.D., Ph.D., Center for Artificial Intelligence in Medicine and Imaging, Stanford University
Why it matters
Self-supervised learning could lead to earlier disease detection and treatment. RETFound provides a promising proof of concept for this.
- RETFound demonstrates a more efficient way to train AI models for medical diagnosis, saving time and money.
- It shows the power of retinal imaging for assessing health conditions.
- RETFound is publicly available for researchers globally to refine and adapt, accelerating innovation.
- The model automates and enhances diagnosis, complementing clinicians, catching issues they may miss, and increasing access to expertise.
- It could reduce healthcare costs if fully validated.
How it works
Using self-supervised learning, RETFound was trained on a dataset of 1.6 million retinal scans.
- This method is similar to the development of natural language processing tools like ChatGPT.
- The model learns what a normal, healthy retina should look like by examining millions of unlabeled examples.
- After the foundation training, the model is fine-tuned to identify specific conditions using a small set of labeled images.
The takeaway
Based on RETFound’s success, medical researchers are eager to apply self-supervised learning to other types of medical imaging data.
Reality check: Care must be taken to ensure ethical use of this method (any AI method) and to communicate limitations.
