Machine Learning on Dynamic Functional Connectivity: Promise, Pitfalls, and Interpretations.

Area of Science:

• Neuroimaging
• Machine Learning
• Cognitive Neuroscience

Background:

• Large-scale functional Magnetic Resonance Imaging (fMRI) data offers opportunities to link brain activity to cognition using data-driven methods.
• Current deep learning models for decoding cognitive states from fMRI data show inconsistent performance across different settings.

Purpose of the Study:

• Establish empirical guidelines for designing deep learning models in neuroimaging.
• Evaluate model performance in cognitive task recognition and disease diagnosis.
• Identify limitations and provide selection criteria for machine learning backbones in neuroimaging.

Main Methods:

• Utilized a large dataset of 39,784 fMRI samples from seven databases.
• Conducted comprehensive evaluations and statistical analyses across cognitive and clinical scenarios.
• Applied an attention-based interpretability method to analyze brain activation patterns.

Main Results:

• No single deep learning model universally outperforms others in neuroimaging applications.
• Model effectiveness is contingent upon factors including demographics, task type, and disease stage.
• Identified key limitations and trade-offs of current deep learning models.

Conclusions:

• Model selection for neuroimaging requires careful consideration of specific application factors.
• Findings provide a foundation for developing more robust and interpretable deep learning models in neuroscience.
• Attention-based interpretability reveals task- and disorder-specific brain activation patterns.