Researchers from Harvard-affiliated Mass General Brigham have developed an AI foundation model designed to extract multiple neurological risk signals from routine brain MR scans.

The model, known as BrainIAC (brain imaging adaptive core), was trained on nearly 49,000 brain MR scans and evaluated across seven tasks spanning a range of clinical complexity. These included estimating brain age, predicting dementia risk, identifying brain tumor mutations, and forecasting survival in patients with brain cancer. The model performed as well as or better than more narrowly focused AI systems, particularly when the amount of task-specific training data was limited, according to a study published in Nature Neuroscience.

Unlike many existing medical imaging models that are trained for single purposes and rely on large, annotated data sets, BrainIAC was built using a self-supervised learning approach. This method allows the system to learn general features from unlabeled MR data before being adapted to specific clinical applications. The researchers said this design helps address common challenges in neuroimaging AI, including variation in MR protocols across institutions and specialties such as neurology and oncology.
After pretraining, the model was validated on 48,965 brain MR scans drawn from multiple data sets. The researchers found that BrainIAC generalized across both healthy and pathological images, performing well on simpler tasks like scan classification as well as more complex ones such as detecting tumor mutation status. In comparisons, it outperformed three task-specific AI frameworks across several benchmarks.

Performance advantages were most pronounced in scenarios with scarce labeled data or higher task complexity, conditions that are common in real-world clinical research and care settings. The authors noted that further studies are needed to assess how well the framework extends to other brain imaging modalities and larger, more diverse data sets.

“BrainIAC has the potential to accelerate biomarker discovery, enhance diagnostic tools, and speed the adoption of AI in clinical practice,” said corresponding author Benjamin Kann, associate professor of radiation oncology at Harvard Medical School and a member of the Artificial Intelligence in Medicine Program at Mass General Brigham.

The study received funding from the National Institutes of Health, the National Cancer Institute, and the Botha-Chan Low Grade Glioma Consortium.

Back to HCB News

Artificial Intelligence Homepage