Feature-reduction and semi-simulated data in functional connectivity-based cortical parcellation

Abstract

Recently, resting-state functional magnetic resonance imaging has been used to parcellate the brain into functionally distinct regions based on the information available in functional connectivity maps. However, brain voxels are not independent units and adjacent voxels are always highly correlated, so functional connectivity maps contain redundant information, which not only impairs the computational efficiency during clustering, but also reduces the accuracy of clustering results. The aim of this study was to propose feature-reduction approaches to reduce the redundancy and to develop semi-simulated data with defined ground truth to evaluate these approaches. We proposed a feature-reduction approach based on the Affinity Propagation Algorithm (APA) and compared it with the classic featurereduction approach based on Principal Component Analysis (PCA). We tested the two approaches to the parcellation of both semi-simulated and real seed regions using the K-means algorithm and designed two experiments to evaluate their noiseresistance. We found that all functional connectivity maps (with/without feature reduction) provided correct information for the parcellation of the semisimulated seed region and the computational efficiency was greatly improved by both featurereduction approaches. Meanwhile, the APA-based feature-reduction approach outperformed the PCAbased approach in noise-resistance. The results suggested that functional connectivity maps can provide correct information for cortical parcellation, and feature-reduction does not significantly change the information. Considering the improvement in computational efficiency and the noise-resistance, feature-reduction of functional connectivity maps before cortical parcellation is both feasible and necessary.

Sample Entropy Combined with the K-Means Clustering Algorithm Reveals Six Functional Networks of the Brain

“…The results indicate that the combination of the K-means clustering algorithm and the FC can identify the functional networks of the human brain. The K-means algorithm has been commonly used to parcellate cortical or subcortical regions based on the static FC [ 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 ]. These previous studies along with the present study extend the application of machine learning methods in brain sciences.…”

“…The K-means clustering algorithm is one of the unsupervised learning algorithms [ 29 ]. Since the K-means clustering algorithm can cluster different observations into different clusters in a simple and easy way, it has been widely used in fMRI studies [ 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 ]. For instance, Fan et al used the K-means clustering algorithm to parcellate the thalamus based on the static FC and found that the thalamus could be divided into seven symmetric thalamic clusters [ 36 ].…”

MonkeyCBP: A Toolbox for Connectivity-Based Parcellation of Monkey Brain

“…PCA, as one of the simplest and robust ways to analyze multidimensional data, is a powerful statistical framework for the analysis of tractography-based parcellation (Thiebaut de Schotten et al, 2014;Cerliani et al, 2017) and has become increasingly popular recently in different studies (Markiewicz et al, 2011;Craddock et al, 2012;Leonardi et al, 2013;Tian et al, 2013;Nayal et al, 2014;Smith et al, 2018). Based on the theory of DWI and tractography techniques, the wholebrain connectivity of each voxel is deemed as multivariate dataset in the process of individual brain parcellation.…”

A Data-Driven Method to Reduce the Impact of Region Size on Degree Metrics in Voxel-Wise Functional Brain Networks

“…The prerequisite of this strategy is to define the sizes of functional regions. In general, one way to approach this problem is to refer to anatomical brain atlases or predefined brain parcellations, but these are either inaccurate or inconsistent ( 23 – 26 ). Moreover, there is no consensus on how many functional regions the brain contains.…”