ESTIMATING REPRODUCIBLE FUNCTIONAL NETWORKS ASSOCIATED WITH TASK DYNAMICS USING UNSUPERVISED LSTMS.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Functional magnetic resonance imaging (fMRI) is crucial for understanding brain activity during tasks.
• Identifying reproducible functional brain networks is essential for reliable neuroscience research.
• Current methods may not fully capture the dynamic relationship between brain networks and task performance.

Purpose of the Study:

• To introduce a novel unsupervised method using LSTMs for estimating reproducible functional brain networks.
• To demonstrate that these LSTM-derived networks are more strongly associated with dynamic task activity.
• To improve the characterization of neural correlates underlying specific cognitive tasks.

Main Methods:

• Utilized recurrent neural networks with long short-term memory (LSTMs) for unsupervised learning.
• Trained the LSTM model to generate functional magnetic resonance imaging (fMRI) time-series data.
• Applied the method to fMRI data from a biological motion perception task, comparing it to existing decomposition techniques.

Main Results:

• LSTM-learned functional networks showed stronger associations with task activity and dynamics compared to other methods.
• The identified network associations were more consistently replicated across subjects and datasets.
• Demonstrated enhanced reproducibility in functional network estimation.

Conclusions:

• The proposed LSTM-based approach yields more reproducible functional networks.
• This method offers improved sensitivity in linking brain network dynamics to task paradigms.
• Enhanced reproducibility is key for robustly characterizing neural correlates in fMRI studies.