Group-fused multivariate regression modeling for group-level brain networks

• Published in: Neurocomputing (Amsterdam) Vol. 363; pp. 140 - 148
• Main Authors: Aggarwal, Priya, Gupta, Anubha
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 21.10.2019
• Subjects: [formula omitted] minimization; Functional connectivity; Functional MRI; Multi-view data; Resting-state brain networks; Subspace clustering; Functional connectivity; Subspace clustering; Resting-state brain networks; l1 minimization; Functional MRI; Multi-view data
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2019.06.042

Currently, functional connectivity (FC) is being used widely to characterize functional brain networks. Availability of huge data, collected from multiple subjects, on 1000 Functional Connectome Project (FCP) and Human Connectome Project (HCP), has made it possible to understand the complex organization of functional brain networks, in particular, in the diseased state, thereby, pressing the need to identify brain networks at the group-level. Often used method to address this problem involves averaging of FC of all subjects followed by the identification of group-level functional brain networks. However, averaging does not account for inter-subject variability. In order to identify effective group-level networks, two issues must be addressed, namely, how to obtain group-fused FC and second, how to combine these group-fused FC matrices to identify group-level networks. In this paper, based on a newly proposed objective function, we incorporate information of subjects via group-fused penalty term. We also impose sparsity and denseness penalties on FC coefficients similar to recently introduced Multivariate Vector Regression-based Connectivity (MVRC) and name the proposed formulation as group-fused MVRC (GF-MVRC). It is then extended to extract group-level brain networks by utilizing iterative spectral clustering. This approach not only identifies group-level networks but also identifies the weighting of each subject’s data in explaining group-level networks, i.e., the proposed work relies on determining subject-level weights responsible for group-level networks. Results on 1000 FCP fMRI dataset corroborate the efficacy of the proposed framework. Experimental results indicate that the proposed GF-MVRC algorithm outperforms existing state-of-the-art atlas based parcellation methods.