Reveal Global Interconnected Functional Tree in the Human Brain

1566 

Abstract Submission 

WEI ZHANG¹, Yu Bao²

¹Augusta University, Augusta, GA, ²James Madison University, Harrisonburg, VA

WEI ZHANG  
Augusta University
Augusta, GA

Yu Bao  
James Madison University
Harrisonburg, VA

Functional connectivity networks (FCNs) portray variance in deoxyhemoglobin concentration consequent to spontaneous or task-evoked modulation of neural metabolism in Blood Oxygen Level Dependent (BOLD) functional Magnetic Resonance Imaging (fMRI) [1], [2]. Since deoxyhemoglobin concentration changes in normal and abnormal brain regions are distinctive, FCNs have been increasingly noticeable to be conducted as a clinically actionable biomarker in order to benefit neurological disorders diagnostics [3-5]. Notably, a single or individual FCN might be more dynamic than structural biomarkers and perhaps unreliable which impairs its application to neurological disease diagnostics. But the unique advantage of FCNs is to reflect the earliest functional variation through various functional regions in the brain before significant irreversible structural impairments happen [6]. In the foreseeable future, a concrete biomarker built on FCNs for early and reliable diagnostics of neurological and psychiatric diseases is desirable. Recent works suggested that understanding the communication of FCNs in nervous systems is a crucial goal in neuroscience and communication of multiple FCNs could be reliable [7]. Thus, we plan to propose a global interconnected functional tree (GIFT) in the normal human brains to represent the functional hierarchy and interconnections across all involved FCNs, which integrates functional interconnection and hierarchy across FCNs.

We propose an innovative deep learning framework named Deep Matrix Approximate Nonlinear Decomposition (DEMAND) to discover reproducible GIFT at the individual level. The following Figure 2 describes a computational framework of DEMAND and Deep Independent Component Analysis (DICA) [8] for validation since ICA is one of the most computational approaches to identify FCNs [8]. In general, the fundamental framework of DEMAND to identify FCNs is to decompose input fMRI signal as a product of time series and canonical or meta-FCNs and plus sub-FCNs. Canonical and meta-FCNs are usually revealed at shallow and deeper layers, as shown in Figure 1 (c2) and (e2), respectively. In addition, since sub-FCNs are usually weak/minor patterns, we treat sub-FCNs as background patterns in Figure 1 (e2). Moreover, to identify hierarchical FCNs, we employ identified FCNs at the previous layer as input to continuously reveal high-level FCNs, such as the decomposition of all identified FCNs at previous layer to the product of time series and FCNs adding sub-FCNs, which is homologous to other layer-stack computational frameworks, shown in Figure 1 (h1).

We validate the proposed DEMAND via employing publicly released Multiband Multi-echo (MBME) fMRI data [9], including test and retest fMRI scans from 70 augmented subjects.
In general, a large identifiability indicates that the difference between test and retest is not significant [10]. Therefore, we present the quantitative results to validate the proposed DEMAND with DICA. Considering identified canonical FCNs, the identifiability of DEMAND ranges from 0.40-0.50. On the contrary, identifiability of canonical FCNs extracted via DICA is lower than 0.38. Meanwhile, the identifiability of meta-FCNs derived via DEMAND is between 0.60-0.70, demonstrating a considerable reproducibility on augmented 70 subjects [9]. The identifiability of meta-FCNs identified via DICA is between 0.35-0.56. The identifiability of sub-FCNs revealed via DEMAND varies from 0.28-0.35 which is relatively small compared to canonical and meta-FCNs. Unfortunately, DICA cannot identify consistent sub-FCNs based on 70 subjects. In Fig. 2, we present a group-wise GIFT on 70 subjects identified via DEMAND.

In a word, the superiority of GIFT is to deliver a global connection abstracted as a graph and more reliable architecture to benefit further understanding of brain functional connectivity and early diagnostics of neurological disorders.

Activation (eg. BOLD task-fMRI) ²

Connectivity (eg. functional, effective, structural) ¹

Computing

Data analysis

FUNCTIONAL MRI

Informatics