No labels? No problem!

Scientists at Harvard Medical School and colleagues at Stanford University have developed an AI diagnostic tool that can detect diseases

A Sept.

The team has exposed the code for the model to other researchers to use

Most AI models need to label datasets during "training" so they can learn to correctly identify pathologies

In contrast, the new model is self-supervised, in the sense that it learns more independently and doesn't require manual labeling of data

"We are living in the early days of the next generation of medical AI models capable of performing flexible tasks by learning directly from text," said Pranav Rajpurkar, principal investigator of the study, who is an assistant professor

Rajpurkar adds: "With CheXzero, one only needs to feed a chest X-ray into the model and the corresponding radiation report, and it will know that the images and text in the report should be considered similar – in other words, it will learn to match

The model was "trained" on a publicly available dataset containing more than 377,000 chest X-rays and more than 227,000 corresponding clinical records

In tests, CheXzero successfully identified pathologies that were not explicitly annotated by human clinicians

The researchers say the method could eventually be applied to imaging methods other than X-rays, including CT scans, MRI and echocardiography

"CheXzero shows that the accuracy of complex medical image interpretation no longer needs to rely on large labeled datasets," said

Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learning