Towards generalisable foundation models for brain MRI.

Area of Science:

• Artificial Intelligence
• Neuroimaging
• Medical Imaging

Background:

• Foundation models are advancing medical imaging but often neglect brain MRI's unique structure.
• Existing models typically focus on 2D natural images, limiting their application to 3D neuroimaging data.

Purpose of the Study:

• To introduce BrainFound, a self-supervised foundation model specifically designed for brain MRI analysis.
• To leverage a slice-based learning strategy for efficient and context-aware processing of MRI volumes.

Main Methods:

• Developed BrainFound, a self-supervised foundation model processing MRI volumes as sequences of 2D slices.
• Enabled single-modality and multimodal input integration (T1, T2, FLAIR) for comprehensive structural information.
• Evaluated performance across neurodegenerative disease detection, tumor grading, and brain tissue segmentation tasks.

Main Results:

• BrainFound consistently outperformed supervised and self-supervised baselines on diverse public datasets.
• Demonstrated strong generalization capabilities, particularly in label-scarce and cross-dataset scenarios.
• Showcased the efficacy of slice-based self-supervised learning for scalable brain MRI analysis.

Conclusions:

• Slice-based self-supervised learning offers a scalable and effective approach for brain MRI analysis.
• BrainFound provides a flexible foundation model for neuroimaging research and potential clinical applications.
• The model shows promise for improving diagnostic accuracy and efficiency in various neurological conditions.