Learning robust and task-invariant functional representation from fMRI through Siamese self-supervised learning

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Functional magnetic resonance imaging (fMRI) is crucial for studying brain function but faces challenges like small sample sizes and poor label quality in specific neurological conditions.
• High dimensionality of fMRI data, coupled with small datasets, increases the risk of model overfitting.
• Existing methods for developing fMRI foundation models require substantial computational resources, limiting their accessibility.

Purpose of the Study:

• To introduce BrainSimSiam, a novel, data-efficient, self-supervised representation learning framework for fMRI data.
• To develop robust and generalizable features from fMRI data using a lightweight framework.
• To address the challenges of limited data and high computational costs in neuroimaging research.

Main Methods:

• Implemented a lightweight self-supervised learning framework named BrainSimSiam.
• Utilized positive-only data pairs to learn representations.
• Evaluated the framework's performance on diverse downstream classification and regression tasks.

Main Results:

• BrainSimSiam-learned representations generalized effectively across various downstream tasks.
• The framework outperformed fully supervised baselines.
• Performance approached that of large-scale models, demonstrating its potential for data-limited scenarios.

Conclusions:

• BrainSimSiam provides a computationally feasible and effective approach for learning from fMRI data, especially in data-limited settings.
• The framework's ability to generate generalizable representations holds significant promise for advancing neuroimaging applications.
• The open-source implementation facilitates wider adoption and further research in the field.