Google has introduced Health AI Developer Foundations (HAI-DEF), a public resource for healthcare developers that provides open-weight models to help them build healthcare applications, initially focused on dermatology, radiology and pathology.
Open-weight AI models are a type of black box AI technology that allows developers to apply and fine-tune a model for specific tasks, allowing them to adapt and build upon previous work.
HAI-DEF is publicly available for developers building and implementing AI models. It includes the open-weight models, documentation to assist in the various stages of development, and instructional Colab notebooks, which allow developers to write and execute code.
The models initially available are aimed at supporting the development of medical imaging applications, such as those relating to chest X-rays, skin images and digital pathology.
The CXR Foundation model for chest X-rays was trained on more than 800,000 X-rays. It allows users to perform data-efficient classification and classify specific conditions, among other tasks.
The Derm Foundation for skin images can be used for data-efficient classification, including dermatitis, melanoma or psoriasis, and for understanding which body part is involved.
Path Foundation for digital pathology is an embedding model that can be used for applications like “grading or identifying tumors, classifying tissue or stain type and determining image quality. The embeddings can also be used for similar image search tasks, to find areas within or across slides that resemble each other,” the tech giant wrote in a blog post.
Google says it introduced HAI-DEF because diverse datasets that include various patient populations, protocols or data-acquisition devices must be available for models to include environments different from the data on which they were trained. The company says the models also allow developers to take their ideas from concept to prototype more easily.
“For healthcare to continue to realize its potential, it needs innovation from a diverse set of contributors on a multitude of use-cases, interfaces and business models,” the tech giant wrote.
HAI-DEF allows developers to download and run the models in their environment locally or via the cloud, use them for research or commercial venture applications, and fine-tune them for better performance.
The models are available via Google’s Vertex AI Model Garden and Hugging Face.
“HAI-DEF is just one of the ways we’re enabling the broader ecosystem to build for health, supplementing Open Health Stack and Population Dynamics Foundation Model. We are excited to continue investing in this space, including by adding more models to HAI-DEF and expanding the scope of our notebooks. We look forward to seeing the community build on these resources to realize AI’s potential to transform healthcare and life sciences,” Google wrote.
THE LARGER TREND
Earlier this year, during Google’s Check Up event, the company announced it was expanding its MedLM models to include multimodal modalities, starting with MedLM for Chest X-ray, available in an experimental preview on Google Cloud. The goal of the model was to enable classification of findings, semantic search and more to improve the efficiency of radiologists’ workflows.
Dr. Ivor Horn, director of health equity and product inclusion at Google, also announced the release of a dermatology-focused dataset, dubbed Skin Condition Image Network (SCIN), that includes skin tones from a diverse group of people with different levels of conditions.
In October, Google announced it is licensing its AI model for detecting diabetic retinopathy to healthcare providers and health-tech partners in Thailand and India, two countries the tech giant says have a shortage of eye specialists.
In June, the company announced the creation of Tx-LLM, an LLM for drug discovery and therapeutic development, fine-tuned from PaLM-2, the company’s generative AI technology that uses Google’s LLMs to answer medical questions.
The tech giant’s medical large language model Med-PaLM 2 was released last year, and was found to generate more comprehensive answers to medical questions than its original version Med-PaLM.
Watch Part 2 of this video interview.
