Researchers at Mass General Brigham and Harvard Medical School have developed an AI foundation model called BrainIAC, designed to analyse brain MRI scans and adapt to the detection of a range of neurological and psychiatric conditions. The work was published in Nature Neuroscience.
Unlike most existing AI tools in medical imaging, which are trained to perform a single specific task, BrainIAC was pre-trained on nearly 49,000 brain MRI scans using a self-supervised learning approach. This allowed the model to develop a broad understanding of brain structure and organisation before being fine-tuned for individual conditions, including Alzheimer’s disease, Parkinson’s disease, autism, dementia, brain tumours, and stroke.
In assessments across these conditions, the model demonstrated competitive performance and showed a notable efficiency advantage: in some cases, it required up to ten times less additional training data to match the performance of models built specifically for a single task. This could be particularly relevant for conditions where large labelled datasets are difficult to obtain.
The authors note that millions of brain MRIs are performed annually, but are typically reviewed for a specific clinical purpose. A flexible foundation model could support researchers in drawing broader insights from existing imaging data, potentially aiding earlier detection and risk prediction across multiple conditions.
BrainIAC is open-source and available to the research community for research use, with clinical utility still requiring further validation. The team reports that several external groups have already begun using the model to investigate conditions, including Alzheimer’s disease and traumatic brain injury. Further validation in clinical settings and across more diverse patient populations will be needed to establish its practical utility.
Published in: Medical Express/ Nature Neuroscience
