Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Global Quantification of the Structural Brain ConnectivityAina Frau-Pascual1, Anastasia Yendiki1, Bruce Fischl1,2, Iman Aganj1,2
1 MGH/HST Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Harvard Medical School2 Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology
Abstract
Method: Global structural brain connectivity
SummaryAcknowledgments
Comparison to standard connectivity
Global connectivity is more correlated with the functional connectivity, than DSI Studio metrics are.
A two-tailed paired t-test between the subtraction of global and DTI SL distributions and global and GQI SL distributions revealed
Global - DTI SL statistic of t = 36.97 significance value of p = 10−60
Global - GQI SL Statistic of t = 35.24 Significance value of p = 10−58
We model brain connectivity globally as a circuit.[Chung et al., 2012, Aganj et al. 2014, Chung et al. 2017]
Results: Alzheimer's disease stage prediction
(a) We use a combination of differential Maxwell’s equations and Kirchhoff’s circuit laws, with diffusion tensors computed from dMRI. [Tuch et al., 2001]
(b) We compute the potentials , as depicted for 5 sources i and a fixed sink s. [O'Donnell et al., 2002]
(c) Then, potential maps are superposed to generate voxel-wise conductance maps as shown for 5 voxels i, or ROI-wise conductance maps as shown for 5 ROIs I.
High conductance between two regions indicates a high degree of connectivity.
DATA analysis:
● MR processings: tissue segmentation and cortex parcellation into ROIs using FreeSurfer.● Diffusion MRI processings: reconstruction of the diffusion tensors (DTI) with DSI Studio, streamline (SL)
tractography using DTI and generalized q-sampling imaging (GQI). ● RS-fMRI processings: detrend, bandpass-filter at 0.01-0.08Hz, and smooth with a 6mm FWHM kernel.
We stacked 4 sessions and computed the correlation matrix for the ROIs.
Using the proposed methodology:● one can compute structural connectivity measures that are significantly more correlated with functional
connectivity than by using more standard approaches. This supports the hypothesis on the role of indirect connections in the relationship between functional and structural connectivity.
● one can better distinguish normal and Alzheimer's disease images.
Distribution of correlations across 100 subjects of the WashU-UMN Human Connectome Project data set.
Project supported by: ● BrightFocus Foundation● National Institutes of Health
Contact: [email protected]
conductivity
x Nvoxels
DTI reconstruction with DSI Studio
(a) Inputs
(c) Conductance ,
(b) Potentials
currents
1
0
0.5
1
0
0.5
10 -3
0
+
-
10 -5
While several studies have shown functional connectivity to be correlated with structural connectivity, strong functional connections have also been commonly observed between regions with no direct structural connection.
Is this variance due to the impact of indirect connections, usually not considered?
DATA: 215 subjects of the Alzheimer's Disease Neuroimaging Initiative (ADNI) data set, in different stages:
CN vs AD connectivity
The regions with decreased connectivity of AD with respect to CN are regions that are known to be affected by AD: temporal regions, hippocampus...
Preliminary results: pair-wise classification prediction accuracy of CN and AD
A Wilcoxon signed-rank test between the median connectivity matrices of each disease stage showed that connectivity matrices are significantly different when comparing any two stages of the disease:
all p-values<0.001 after Bonferroni correction.
A one-sample t-test on the element-wise t-values of the different stages confirmed this result:
all p-values<0.001 after Bonferroni correction, except for CN / SMC groups
Comparison of disease stages
MOTIVATION: Brain connectivity relies on a pipeline with many parameters, and the connectivity results vary depending on the decisions made in this pipeline. The field could benefit from a more robust connectivity measure.
GOAL: Finding a new structural connectivity measure that considers all possible pathways, direct and indirect, and evaluating the relationship of this measure with functional connectivity and its role in the classification of diseased and healthy populations.
AdaBoost with base estimator decision tree classifier (n=20) of 5 PCA components:
● Proposed conductance method:
80.00% prediction accuracy
● GQI-SL counting tracts that cross an ROI, normalized by the median length:
68.95% prediction accuracy
Cognitively Normal (CN): 47 subjectsSignificant Memory Concern (SMC): 56 subjectsEarly Mild Cognitive Impairment (EMCI): 34 subjectsLate Mild Cognitive Impairment (LMCI): 49 subjectsAlzheimer's disease (AD): 31 subjects
subtraction of median connectivities of AD and CN groups
tractography: recon DTI, GQI, streamline
connectivity measure: tract count, normalized tract count by median length...
global connectivity
N ROI
N ROI
1 ROI
conductance connectivity matrix
standard connectivity
conductance model
Gray matter segmentation with FreeSurfer
DTI reconstruction with DSI Studio
circuit solution with sources in 1 ROI
tracts end in ROI
tracts cross ROI
count norm. count
Results: structural vs. functional connectivity