Artificial intelligence (AI) thrives on data. While chatbots and self-driving cars easily find their training data online, one would likely assume that finding equivalent data for medical AI online presents a much tougher challenge. Yet, Stanford Medicine researchers found an unexpected solution: Twitter, now rebranded as ‘X’.
Tapping into X’s rich database, researchers trained an AI algorithm using over 200,000 pathology images. This algorithm is capable of interpreting varied medical conditions like melanoma, breast cancer, and even parasitic infections. Though it doesn’t diagnose, it’s a goldmine for clinicians and students seeking reference images.
James Zou, Ph.D., highlighted its primary use: “It helps human pathologists find similar cases for reference.”
X’s Unexpected Role
While X might seem an unlikely medical resource, it’s been a hidden treasure trove. With the pathology community’s 32 specialized hashtags, pathologists have been actively sharing crucial medical images and engaging in discussions.
These posts typically pair an image with a brief description. This combination proved perfect for training the algorithm, allowing it to bridge visual data with text.
OpenPath: A Pathology Database
The research team meticulously curated relevant posts from 2006-2022, finally collating over 200,000 image-text pairs. This collection, named OpenPath, is now a giant in public datasets of human-annotated pathology images.
After training on OpenPath, the AI model, dubbed PLIP, emerged. It uses advanced techniques to match images with text descriptions, acting like a specialized Google Image search for pathologists.
PLIP: A Pathologist’s AI Assistant
When tested, PLIP impressively outscored existing models. But Zou underscores its role as a supporter, not a replacement, for human pathologists.
With PLIP’s success, there’s potential to expand this approach to fields like radiology and dermatology. As data keeps pouring into PLIP, Zou confirms, “It only gets better with more data.”
