Constrained Sparse Functional Connectivity Networks for MCI Classification

Abstract

Abstract. Mild cognitive impairment (MCI) is difficult to diagnose due to its subtlety. Recent emergence of advanced network analysis techniques utilizing resting-state functional Magnetic Resonance Imaging (rs-fMRI) has made the understanding of neurological disorders more comprehensively at a whole-brain connectivity level. However, inferring effective brain connectivity from fMRI data is a challenging task, particularly when the ultimate goal is to obtain good control-patient classification performance. Incorporating sparsity into connectivity modeling can potentially produce results that are biologically more meaningful since most biologically networks are formed by a relatively few number of connections. However, this constraint, when applied at an individual level, will degrade classification performance due to inter-subject variability. To address this problem, we consider a constrained sparse linear regression model associated with the least absolute shrinkage and selection operator (LASSO). Specifically, we introduced sparsity into brain connectivity via l1-norm penalization, and ensured consistent non-zero connections across subjects via l2-norm penalization. Our results demonstrate that the constrained sparse network gives better classification performance than the conventional correlation-based network, indicating its greater sensitivity to early stage brain pathologies.

Functional Brain Network Classification With Compact Representation of SICE Matrices

“…Note that because SLR matrices are not necessarily SPD, the SPD kernels cannot be applied. Therefore, no classification result is reported in the row of "SLR" in Table II. 2) Classification using the compact representation: In this experiment, we compare the classification performance of the compact representation obtained by the proposed SPDkernel PCA, linear PCA and the method computing local clustering coefficient (LCC) [17]. LCC, as a measure of local neighborhood connectivity for a node, is defined as the ratio of the number of existing edges between the neighbors of the node and the number of potential connections between these neighbors [42].…”

“…The normalized brain images are warped into automatic anatomical labeling (AAL) [41] atlas to obtain 90 ROIs as nodes. By following common practice [15], [16], [17], the ROI mean time series are extracted by averaging the time series from all voxels within each ROI and then bandpass filtered to obtain multiple sub-bands as in [17].…”

“…For comparison, constrained sparse linear regression (SLR) [17] is also used to learn functional connectivity networks with the same setting. Functional connectivity networks constructed by SICE and SLR are called "SICE matrices" and "SLR matrices" respectively.…”

“…independent component analysis [11], [12], and clustering-based methods [13], [14]. Given a set of network nodes, the functional connectivity between two nodes is conventionally measured by the correlation coefficient of time series associated with the two nodes (e.g., the averaged time series from all voxels within a node) [15], [16], [17], and the brain network is then represented by a correlation matrix.…”

“…This usually leads to poor performance of classification. An alternative approach is to summarize a d × d SICE matrix into lower dimensional graphical features such as local clustering coefficient (LCC) [17] or hubs [23]. Nevertheless, these approaches have the risk of losing useful information contained in the SICE matrices.…”