Neighborhood structure-guided brain functional networks estimation for mild cognitive impairment identification

Area of Science:

• Neuroimaging and computational neuroscience.
• Development of advanced analytical techniques for brain functional network construction.

Background:

• Functional magnetic resonance imaging (fMRI) and Pearson's correlation (PC) are crucial for brain disease diagnostics, constructing brain functional networks (BFNs).
• Existing methods using PC often result in dense BFNs, violating biological principles, and thresholding or L1-norm regularization neglect spatial information.
• Regions of interest (ROIs) with closer distances typically exhibit higher connectivity due to wiring cost principles.

Purpose of the Study:

• To propose a novel neighborhood structure-guided method for estimating sparse and biologically plausible BFNs.
• To incorporate spatial neighborhood information into BFN construction, addressing limitations of current PC-based approaches.
• To evaluate the efficacy of the proposed method in distinguishing individuals with mild cognitive impairment (MCI) from healthy controls.

Main Methods:

• Calculate Euclidean distances between ROIs and sort them.
• Employ K-nearest neighbor (KNN) to identify the K closest neighbors for each ROI.
• Construct a global topology adjacency matrix based on these K-neighbor relationships, then use PC to calculate correlations for connected ROIs, generating the BFN.

Main Results:

• The proposed neighborhood structure-guided method (K-nearest neighbor-Pearson's correlation, K-PC) achieved better classification performance in distinguishing MCI from healthy individuals compared to baseline methods.
• K-PC demonstrated an advantage over commonly used deep learning time series methods.
• The method generates sparser and potentially more biologically relevant BFNs by considering spatial proximity.

Conclusions:

• The neighborhood structure-guided BFN estimation method effectively improves classification accuracy for neurological conditions like MCI.
• Incorporating spatial neighborhood information enhances the biological relevance and diagnostic utility of BFNs derived from fMRI.
• This approach offers a promising alternative to existing methods, including deep learning, for brain network analysis.