State-space model with deep learning for functional dynamics estimation in resting-state fMRI.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Resting-state functional Magnetic Resonance Imaging (rs-fMRI) reveals dynamic interactions within large-scale brain networks.
• These functional dynamics are crucial for understanding brain function and dysfunction, presenting a challenge for current network modeling.

Purpose of the Study:

• To propose a novel methodological architecture combining deep learning and state-space modeling for analyzing time-varying functional networks in rs-fMRI.
• To apply this architecture to diagnose Mild Cognitive Impairment (MCI) by capturing functional dynamics.

Main Methods:

• A Deep Auto-Encoder (DAE) was developed to identify hierarchical, non-linear functional relationships and transform regional features into an embedding space.
• A Hidden Markov Model (HMM) was employed to infer dynamic characteristics of functional networks from the embedded features via hidden states.
• A generative model using HMM was built to estimate the likelihood of rs-fMRI features belonging to MCI or normal control groups.

Main Results:

• The proposed method demonstrated effectiveness in experiments on two datasets, outperforming state-of-the-art approaches.
• Analysis of learned functional networks by DAE and estimated connectivities by HMM provided insights into brain dynamics.
• Graph-theoretic analysis further investigated the estimated functional connectivities.

Conclusions:

• The combined DAE and HMM approach offers a powerful tool for modeling functional dynamics in rs-fMRI.
• This methodology shows significant potential for improving the accuracy of MCI diagnosis through the analysis of dynamic brain network alterations.