Difference between revisions of "Brain Network Analysis"

Revision as of 08:51, 6 May 2015

The aim of this project is to explore the networks properties (global and local) over brain activity networks.

Data Sets

Resting State Networks

Experimental evidence obtained by using fMRI in resting state (fMRI-RS) suggests that the brain may be organized on regions of cognitive relevance that are functionally connected, in the so called resting state networks (RSNs).

Resting state networks in a healthy patient

The term functional connectivity refers to the connectivity between brain regions that exhibit synchronized brain activity. A RSN encompasses regions related to lower and higher brain functions. RSNs can be considered as high level cerebral entities that may account for cognitive and behavioural information. At least ten of these cerebral entities have been consistently identified in control subjects, namely, default network, left and right executive networks, auditory, salience, sensorimotor, and three visual networks cerebellum (lateral, medial and occipital), as observed in Figure 1. The study of the dynamics associated to these RSNs, for healthy and altered brain conditions, is a highly active research area in neurosciences.

Participants

Data from 76 subjects were used for this study: 27 healthy controls (14 women, mean age 47 +- 16 years), 24 patients in minimal conscious state and 25 with vegetative state/unresponsive wakefulness syndrome (20 women, mean age 50 +- 18 years). All patients were clinically examined using the French version of the Coma Recovery Scale Revised (CRS-R). Exclusion criteria were contra-indication for MRI (e.g., presence of ferromagnetic aneurysm clips, pace-makers), MRI acquisition under sedation or anaesthesia and large focal brain damage (>50 % of total brain volume). The study was approved by the Ethics Committee of the Medical School of the University of Liege in Belgium. Written informed consent to participate in the study was obtained from all patients or legal surrogates of the patients.

Data acquisitions and preprocessing

The healthy volunteers was instructed to close their eyes, relax without falling a sleep and refrain from any structured thinking (e.g., counting, singing etc.). The same instructions were given to patients but due to their cognitive and physical impairments, we could not fully control for a prolonged eye-closed yet awake scanning session. For each subject, fMRI resting data were acquired in a 3T scanner (Siemens medical Solution in Erlangen, Germany). Three hundred fMRI volumes multislice T2-weighted functional images were captured (32 slices; voxel size: 3x3x3;mm3; matrix size 64x64x32; repetition time = 2000 ms; echo time = 30 ms; ip angle = 78°; field of view = 192x192 mm2). The three initial volumes were discarded to avoid T1 saturation efects. An structural T1 image was also acquired for anatomical reference.

fMRI data was processed using SPM8. Preprocessing includes: realignment, coregistration of functional onto structural data, segmentation of structural data, normalization into MNI space and spatial smoothing with a Gaussian kernel of 8 mm. Large head motions were corrected using ArtRepair.

Methods and Measurements

Functional Network Connectivity

1. Spatial Independent Component Analysis: The first step for the RSN identification was the fMRI signal decomposition into sources of neuronal/physiological origin. For this task, we used ICA, which aims to decompose the signal into a set of statistically independent components (ICs) of brain activity. Because in the fMRI data the spatial dimension is much greater than temporal one, we used spatial ICA (sICA), which decompose the signal into maximally independent spatial maps. In sICA each spatial map (source) have an associated time course, which corresponds to the common dynamic exhibit by this component. These RSN time courses were subsequently used for all the FNC computations. For the ICA decomposition we used 30 components and the infomax algorithm as implemented in GroupICA toolbox 3 .
2. RSNs Identification: After the ICA decomposition, the different RSNs were identified at individual level. The common approach for this task is the group level identification. In this method, the fMRI data of whole population is concatenated along the temporal dimension. Later, sICA is applied to identify the sources of brain activity at the group level. Following, each RSN is manually identified. Finally, individual time courses are extracted for each RSN by applying a dual regression (back-reconstruction) onto the original subject data. This approach is based on a homogeneity assumption of the fMRI dynamic across the whole population. Nevertheless, in severely affected brains, this condition may be not valid. After the RSN spatial map identification, a machine learning based labeling method was applied to discriminate between IC of “neuronal”or “artifactual”origin. In particular, a binary classification method based on support vector machines and an spatio-temporal feature vector for description each IC was used.
3. Time series interaction measurement:
   • Pearson Correlation: PC captures lineal relationships among time courses.
   • Normalized Mutual Information:
   • Distance Correlation: DC aims to measure dependencies between two random variables X and Y with finite moments in arbitrary dimension, not necessarily of equal dimensions. For the FNC computations we assumed that two RSN time series X and Y provide the n observations of the joint distribution characteristic of the RSN temporal dynamics. Prior to the DC computations, the RSN time courses were filtered thought a bandpass Butterworth filter with cut-off frequencies set at 0.05 Hz and 0.1 Hz.

Individual Subject FNC analysis in two subjects. Thickness corresponds to connectivity strength

Complex Network Measurements

In order to explore networks, a set of different networks measurements were applied to each red. The measurements were reported by "Mikail Rubinov and Olaf Sporns" in Complex networks measures of brain connectivity: Uses and interpretations. Here a short description of the measures used in this exploratory exercise:

• degree: The degree of an individual node is equal to the number of links connected to that node.
• degree distribution: The degrees of al nodes in the network.
• clustering coefficient: The fraction of triangles around an individual node. It is equivalent to the fraction of neighbour nodes that are also neighbour of each other.
• local efficiency: It is the global efficiency computed on node neighborhoods, and it is related to the clustering coefficient.
• global efficiency: Average inverse shortest path length. It is meaningfully computed on disconnected networks.
• distance: Distance matrix (Dijkstra's algorithm). The input matrix must be a mapping from weight to distance (usually weight inversion).
• characteristic path: the average shortest path length between all pairs of nodes in the networks.
• eigenvector centrality: Eigenvector centrality is a self-referential measure of centrality -- nodes have high eigenvector centrality if they connect to other nodes that have high eigenvector centrality.
• outreach:
• averaged shortest path: Same Characteristic path. The global efficiency is the average inverse shortest path length in the network. The node eccentricity is the maximal shortest path length between a node and any other node. The radius is the minimum eccentricity and the diameter is the maximum eccentricity.
• modularity: The modularity is a statistic that quantifies the degree to which the network may be subdivided into such clearly delineated groups.
• transitivity: The transitivity is the ratio of triangles to triplets in the network and is an alternative to the clustering coefficient.
• Rich club: The rich club coefficient at level k is the fraction of edges that connect nodes of degree k or higher out of the maximum number of edges that such nodes might share.

brain connectivity

Results

In order to compute the degree, strength, clustering, eigenvector centrality, distance, characteristic path and global efficiency for all population, the measurements were made by subject; then, the mean and standard deviation were calculated from population measurements.

The next results were obtained using Distance Correlation for computing the correlation of brain networks.

Degree

Strength

Clustering

Local Efficiency

Results without some networks

From the literature is known that some networks were affected by the pathology. These networks are: Executive Control Left, Executive Control Right and Visual Media. Then, the measurements were recomputed with the data without consider these three networks.

Degree

Strength

Clustering