A generalizable foundation model for analysis of human brain MRI

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Informatics

Background:

• Artificial intelligence (AI) in brain magnetic resonance imaging (MRI) shows promise for neurological disease diagnosis, prognosis, and treatment.
• Current AI models face limitations due to restricted training data and poor generalization across diverse patient populations and tasks.
• Foundation models, utilizing self-supervised learning, pretraining, and adaptation, offer a potential solution to these challenges.

Purpose of the Study:

• To introduce Brain Imaging Adaptive Core (BrainIAC), a foundation model for generalized representation learning from unlabeled brain MRI data.
• To establish a core model adaptable for various downstream applications in neuroimaging analysis.
• To overcome limitations of traditional AI models in brain MRI analysis.

Main Methods:

• Developed BrainIAC, a foundation model leveraging self-supervised learning on a large dataset of unlabeled brain MRIs.
• Trained and validated BrainIAC on 48,965 brain MRIs across multiple tasks.
• Employed targeted adaptation strategies for downstream applications.

Main Results:

• BrainIAC demonstrated superior performance compared to localized supervised training and other pretrained models.
• The model excelled in low-data, few-shot learning scenarios and high-difficulty prediction tasks.
• Performance improvements were significant in settings where traditional methods were previously infeasible.

Conclusions:

• BrainIAC offers a powerful, generalizable foundation model for brain MRI analysis, addressing data scarcity and task specificity issues.
• The model facilitates improved biomarker discovery and accelerates the clinical translation of AI in neurology.
• BrainIAC's integration into imaging pipelines and multimodal frameworks holds significant potential for advancing neurological disease care.