HKUST Develops Novel AI Framework Enabling Efficient Collaboration between Generalist and Specialist Models for Disease Diagnosis
Prof. Hao CHEN, Assistant Professor of the Department of Computer Science and Engineering and the Department of Chemical and Biological Engineering and Director of the Collaborative Center for Medical and Engineering Innovation, has led an international research team to develop a groundbreaking AI framework called Generalist–Specialist Collaboration (GSCo). This system allows a broad-knowledge "generalist" AI and highly focused "specialist" AIs to work together as a team, combining the generalist's adaptability with the specialists' precision.
By having the generalist AI act as a smart decision-maker that evaluates advice from the specialist AIs, the combined system significantly outperforms standalone models in analyzing medical scans, answering clinical questions, and drafting radiology reports.
Crucially, this new approach solves major real-world hurdles for healthcare:
- Guards Patient Privacy: Specialist models can be trained directly inside individual hospitals using local data, ensuring sensitive patient records never have to leave the institution.
- Drastically Lowers Costs: Adapting the system to a new medical task is up to 100 times cheaper than updating a massive, standalone AI model.
- Reduces Tech Barriers: The specialist models are lightweight and can be trained using just a single standard computer graphics card, making advanced AI accessible even to clinics with limited funding and computing power.
The research titled "Towards generalizable AI in medicine via Generalist–Specialist Collaboration," has been published in the prestigious journal Nature Biomedical Engineering, and the team has made their code publicly available to help speed up the adoption of safe, affordable AI assistants in everyday clinics. Details are available at https://github.com/sunanhe/MedDr.
(This news was originally published by the HKUST Global Engagement and Communications Office here.)
Related Links
