The big companies brokering artificial intelligence (AI) tools have been actively discussing whether the code and data behind AI models should be available to everyone (open models) or if access should be limited (closed models).
Google has leaned open in its AI strategy, making the tech behind its flagship Gemini system widely available earlier this year—and it’s only doubled down.
On November 25, the company announced it would open its Health AI Developer Foundation models (HAI-DEF) for three types of medical image categories: chest X-ray, dermatology, and pathology.
AI could help in clinical settings with tasks like offering personalized diagnosis or prognosis suggestions or assisting with triage. But it’s hard to build AI models from scratch because the amount of health-specific data needed to train and test them usually cannot be supplied by just one institution. This is why many models tend to perform well in the lab, but fail in clinical settings.
Google has accrued a diverse range of medical imaging data and other clinically relevant information from its partners and put it into a basic working model.
For example, the dermatology foundation model has been trained on pictures of skin conditions and related text information drawn from the internet, according to a technical paper published by Google researchers in preprint journal arXiv. In research use cases, the authors wrote, Google’s pathology foundation model helped University College London researchers distinguish different types of sarcomas.
“The intention behind this is to allow developers, essentially, to accelerate the development of AI solutions that can be built on top of this,” Shravya Shetty, engineering director at Google, told Healthcare Brew.
Organizing healthcare-specific data. This package has been the culmination of nearly a decade of work, Shetty said. Google Research has been testing and developing AI tools with partners across healthcare organizations and pharmaceutical companies like Apollo Radiology in India, the Maryland-based research hospital National Institutes of Health Clinical Center, and health tech company DeepHealth.
Navigate the healthcare industry
Healthcare Brew covers pharmaceutical developments, health startups, the latest tech, and how it impacts hospitals and providers to keep administrators and providers informed.
Since starting, the models have evolved from text- or image-only identification deep learning models to become more sophisticated and multimodal.
These models were initially available around 2022 as APIs, or application programming interfaces, which allowed for research access (groups can submit API requests to Google, which has to approve their permission to use the model). But research-only access created “bottlenecks” that prevented people from maximizing the models’ utility, Shetty said.
Open-weight models, on the other hand, would solve the constraint and limits around access, she added, since they can provide developers with more transparency and control, especially over sensitive information, like health data, that might be fed to the model.
“This means that you can take the models as is and tune it on your data,” Shetty said. “People are free to build solutions on top of it, do their own validation, and deploy it in the manner of their choosing.”
Plus, the models are not disease-specific. For example, Shetty said, the chest X-ray model can be used to identify thousands of conditions, from tuberculosis to rare lung diseases.
“I don’t think that there will ever be a single model that will satisfy all the various nuances in healthcare. There’s so many conditions, so much specialization for individuals,” she said. “The only way it scales—and for it to reach its potential—is if the industry as a whole moves in this direction.”