Boston University MRI model hits 83% Dice score with 10 scans

Boston University researchers built a transformer AI model for brain MRI analysis that hits an 83 percent Dice score with fewer than 10 labeled scans, potentially cutting diagnostic delays for stroke and dementia patients.

"In many cases, hospitals have access to imaging data, but not enough expert-labeled examples to train large models from scratch," Xin Zhang, distinguished professor of engineering at BU's College of Engineering, said. "We wanted to build a system that can learn from what is available and still perform well when labeled data are limited."

The team trained a masked autoencoder on 31 million unlabeled, de-identified MRI images drawn predominantly from large public data sets. The model can be fine-tuned for clinical tasks including image restoration, sequence detection, skull stripping, and brain region segmentation. In tests, it outperformed rival research groups at identifying the boundary between skull and brain tissue, and earned a Dice score — a metric used in evaluating machine learning performance — above 83 percent even when fine-tuned with fewer than 10 annotated scans.

The technology could cut MRI exam times from 20-30 minutes to 12-15 minutes, reducing patient motion artifacts and enabling faster diagnosis. Chad Farris, assistant professor of radiology at BU's Chobanian & Avedisian School of Medicine and a neuroradiologist at Boston Medical Center, said early detection could be clinically significant for conditions like stroke where "even an hour or two could make a big difference."

Why pretrained models matter for medical imaging

The approach addresses two barriers that have kept AI out of clinical radiology. First, current methods require separate models trained from scratch for every disease or condition — developers are building independent algorithms for hemorrhage detection, aneurysm detection, and other tasks, Farris said. Second, training such models requires thousands of manually annotated images, a process that "takes forever" because neuroradiologists must label each scan by hand, Zhang said.

"You can't get 10,000 annotated brain MRIs," Zhang said. "It's just not practical."

The masked autoencoder approach sidesteps both problems. By randomly blanking out patches of images during pretraining, the model learns general patterns in brain anatomy without needing labels. Clinics can then adapt it to specific tasks with a handful of verified examples.

The findings, published in Frontiers in Artificial Intelligence, come as patients increasingly turn to general-purpose AI chatbots for medical interpretation. A 2026 KFF poll found that 33 percent of American adults use AI chatbots for health advice, and 19 percent use AI specifically to interpret lab results or medical tests. OpenAI said more than 300 million people ask health-related questions on ChatGPT each week.

But general-purpose chatbots remain unreliable for medical imaging. Mengyu Li, a mechanical engineering doctoral student in Zhang's lab, tested one commercial chatbot that confused a knee with a spine. "The response may sound correct, but it might actually be wrong," Li said. "That's why, to this day, there is still no FDA-approved agentic AI for analyzing MRIs."

The competitive field and investment angle

The BU model enters a crowded field of medical AI efforts. Epic Systems, the electronic medical-records provider, has built an AI assistant into its software. Anthropic's Claude can connect to some users' medical records. OpenAI rolled out a health feature in July that links with users' medical records and wearables data.

Yet none of these general-purpose tools has received FDA approval for MRI analysis. The BU team's approach — pretraining on unlabeled data and fine-tuning with minimal labels — could give hospitals a path to deploy AI without the prohibitive cost of manual annotation.

For investors, the implications span multiple sectors. Medical imaging AI companies could see their data advantage shrink as pretrained models reduce the need for large annotated data sets. Radiology departments at institutions like Boston Medical Center could see faster turnaround times and lower costs. The technology could eventually extend beyond brain imaging to other organs, Farris said, calling it "a stepping stone for building programs able to identify diseases earlier and more accurately than we're currently able to do with imaging alone."

This article is for informational purposes only and does not constitute investment advice.