VarCoNet: A Variability-Aware Self-Supervised Framework for Functional Connectome Extraction From Resting-State fMRI

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Interindividual variability in brain function is crucial for precision medicine.
• Existing methods often treat functional variability as noise, overlooking its potential.
• Robust functional connectome (FC) extraction from resting-state fMRI (rs-fMRI) is essential for understanding brain function.

Purpose of the Study:

• Introduce VarCoNet, a self-supervised framework for enhanced FC extraction from rs-fMRI data.
• Leverage functional interindividual variability as meaningful data for brain function encoding.
• Develop a versatile framework applicable to downstream tasks like subject fingerprinting and clinical diagnosis.

Main Methods:

• VarCoNet employs self-supervised contrastive learning with a novel rs-fMRI signal segmentation augmentation strategy.
• The framework integrates a 1D-CNN with a Transformer encoder for advanced time-series analysis.
• Bayesian hyperparameter optimization enhances the model's robustness.

Main Results:

• VarCoNet achieved up to 98% accuracy in subject fingerprinting using Human Connectome Project data.
• For autism spectrum disorder (ASD) classification, VarCoNet reached an AUC of 72.6% on ABIDE I and II datasets.
• Extensive comparisons against 13 deep learning methods demonstrated VarCoNet's superior performance, robustness, and generalizability.

Conclusions:

• VarCoNet offers a versatile and robust framework for FC analysis in rs-fMRI data.
• The framework effectively utilizes functional interindividual variability for improved brain function encoding.
• VarCoNet shows significant potential for applications in precision medicine and neurological disorder research.