When Mengyu Li tested a popular chatbot on a medical image, it confidently mistook a knee for a spine. The mistake was funny in a research paper, but in a real hospital it could be deadly. That gap between what AI can do in a chat box and what doctors need at a hospital bed is exactly what a team at Boston University is trying to close.
Xin Zhang, a distinguished professor of engineering at BU's Laboratory for Microsystems Technology, and her colleagues have built an AI tool that can analyze brain MRI scans even when a clinic has only a handful of labeled examples to work with. Normally, teaching a computer to read brain scans requires thousands of verified examples. Many hospitals simply do not have that volume of carefully annotated data. Zhang's team designed their system to learn from whatever is available and still produce reliable results.
"In many cases, hospitals have access to imaging data, but not enough expert-labeled examples to train large models from scratch," Zhang said. "We wanted to build a system that can learn from what is available and still perform well when labeled data are limited."
The tool is built on a transformer — the same technology that powers chatbots like ChatGPT. But unlike a chatbot that can invent answers that sound plausible but are completely wrong, this system learns to recognize patterns in actual brain images. It can spot local details within a single scan while also building a larger picture of how healthy brains look on average. That two-level understanding lets it flag when something seems off, even without seeing thousands of examples of that specific problem first.
The team calls it a kind of Swiss Army knife for MRI analysis. One day it might detect lesions; the next, it could flag early signs of dementia. Chad Farris, an assistant professor of radiology at BU Chobanian & Avedisian School of Medicine and a neuroradiologist at Boston Medical Center, says the technology could be transformative for patients.
"Early detection could be clinically significant," Farris said. "When you get an MRI, it might take a few days or a week for it to get read. But there could be an acute reading, like a stroke, that you'd want to know about right away. Even an hour or two could make a big difference." In some cases, Farris explained, doctors cannot confirm a diagnosis until after a patient passes away — a reality that early screening might one day change.
Currently, no AI system has received FDA approval for independently analyzing MRI scans. Zhang's team published their findings in Frontiers in Artificial Intelligence. With further development, their tool could give hospitals — especially smaller ones with limited data — a powerful new way to catch serious conditions before they become emergencies.