A new AI foundation model has been developed that can accurately extract multiple disease risk signals from routine brain MRIs, including: estimating a person’s “brain age”; predicting dementia risk; detecting brain tumor mutations; and predicting survival from brain cancer, according to investigators from Harvard-affiliated Mass General Brigham.

The model, a brain imaging adaptive core called BrainIAC, was trained on nearly 49,000 brain MRI scans. The tool outperformed other, more task-specific AI models, and was especially efficient when limited training data were available. 

Results are published in Nature Neuroscience.

According to researchers, despite recent advances in medical AI approaches, there is a lack of publicly available models that focus on broad, brain MRI analysis. Most conventional frameworks perform specific tasks and require extensive training with large, annotated datasets that can be hard to obtain. Furthermore, brain MRI images from different institutions can vary in appearance and based on their intended applications (such as in neurology versus oncology care), making it challenging for AI frameworks to learn similar information from them.

To address these limitations, BrainIAC uses a method called self-supervised learning to identify inherent features from unlabeled datasets, which can then be adapted to a range of applications. After pretraining the framework on multiple brain MRI imaging datasets, the researchers validated its performance on 48,965 diverse brain MRI scans across seven distinct tasks of varying clinical complexity.

Researchers found that BrainIAC could successfully generalize its learnings across healthy and abnormal images and subsequently apply them to both relatively straightforward tasks, such as classifying MRI scan types, and very challenging tasks, such as detecting brain tumor mutation types. The model also outperformed three more conventional, task-specific AI frameworks at these applications and others.

The authors note that BrainIAC was especially good at predicting outcomes when training data was scarce or task complexity was high, suggesting that the model could adapt well to real-world settings where annotated medical datasets are not always readily available. Further research is needed to test this framework on additional brain imaging methods and larger datasets.

“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 of the Artificial Intelligence in Medicine (AIM) Program at MGB and associate professor of radiation oncology at Harvard Medical School. “Integrating BrainIAC into imaging protocols could help clinicians better personalize and improve patient care.”

This study was supported in part by the National Institutes of Health/the National Cancer Institute and Botha-Chan Low Grade Glioma Consortium.