Towards a general-purpose foundation model for functional MRI analysis.

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Functional magnetic resonance imaging (fMRI) is vital for brain research and diagnosing neurological conditions.
• Current fMRI analysis methods face challenges in reproducibility and transferability due to complex pipelines and task-specific models.

Purpose of the Study:

• To introduce NeuroSTORM (Neuroimaging Foundation Model with Spatial-Temporal Optimized and Representation Modelling), a novel foundation model for fMRI analysis.
• To enable learning of generalizable representations from fMRI data for diverse downstream applications.

Main Methods:

• NeuroSTORM utilizes a foundation model approach, pretraining on a large dataset of 28.65 million fMRI frames from over 50,000 participants (ages 5-100).
• The model incorporates an efficient spatiotemporal modeling design and lightweight task adaptation for scalable pretraining and rapid transfer.
• Direct learning from four-dimensional (4D) fMRI volumes is employed to capture complex brain dynamics.

Main Results:

• NeuroSTORM consistently outperformed existing methods across five downstream tasks: demographic prediction, phenotype prediction, disease diagnosis, re-identification, and state classification.
• On clinical cohorts with 17 diagnoses, NeuroSTORM achieved superior diagnostic performance.
• The model demonstrated predictive capabilities for psychological and cognitive phenotypes.

Conclusions:

• NeuroSTORM offers a standardized foundation model for reproducible and transferable fMRI analysis.
• The model's ability to generalize across diverse tasks and clinical data highlights its potential impact on neuroscience and clinical practice.