PowerPoint Presentation
Computational approaches for mapping the human connectome R.
Cameron Craddock, PhDDirector, Computational Neuroimaging LabNathan
S. Kline Institute for Psychiatric ResearchDirector of Imaging,
Center for the Developing BrainChild Mind Institute
March 30, 2016
1
Functional Magnetic Resonance Imaging (fMRI)An fMRI time-course
is formed by the rapid acquisition of MR images which are sensitive
to the blood oxygen level dependent (BOLD) contrast
Hemoglobin, the protein which transports oxygen in blood,
contains four heme molecules, each with an atom of
ironDeoxy-hemoglobin is paramagnetic and creates a magnetic
gradient that dephases the MRI signalOxy-hemoglobin is diamagnetic
and does not affect the MRI signal
Hemodynamic ResponseInitially neuronal oxygen consumption
increases the amount of deoxy-hemoglobin and the MR signal
decreasesBlood flow increases, bringing more oxygenated blood to
the area than is required, resulting in a signal increaseWhen
neuronal activity ceases, the signal returns to baseline after a
brief undershoot
Brain Mapping with fMRI
Resting State Functional Connectivity
Biswal et al. MRM 1995Intrinsic activity is ongoing neural and
metabolic activity which is not directly associated with subjects
performance of a task-Raichle TICS 2010
Intrinsic Connectivity Networks
De Luca et al, 2006Independent Component Analysis
Kelly et al, 2007Functional Connectivity Seeding
Functional Connectivity AnalysisData are preprocessedIndividual
level FC maps are generated by ROI correlation with rest of brain,
ICA, cross-correlation between several ROIs, or other methodFC maps
are compared between groups feature-by-feature (voxel-by-voxel)
using t-tests
Data Driven ROI AtlasCraddock et al. Human Brain Mapping
2012.
Functional Connectivity Graphs
Courtesy of Dr. Xinian Zuo
The Human ConnectomeThe sum total of all of the brains
connectionsStructural connections: synapses and fibersDiffusion MRI
Functional connections: synchronized physiological activityResting
state functional MRI
Nodes are brain areasEdges are connections
Craddock et al. Nature Methods, 2013.
Discovery science of human brain functionCharacterizing
inter-individual variation in connectomes (Kelly et al. 2012)
Identifying biomarkers of disease state, severity, and prognosis
(Craddock 2009)
Re-defining mental health in terms of neurophenotypes, e.g. RDOC
(Castellanos 2013)
MVPA ClassificationGiven a training set of observations and
corresponding labels (categorical) objective is to find a linear
hyperplane capable of discriminating observations For linearly
separable data there are an infinite number of hyper-planes that
meet this requirementSVM finds the unique hyperplane that maximizes
the perpendicular distance from the hyperplane to the nearest
observations of a class (margin)
Train vs. TestTraining or learning is the process of optimizing
MVPA modelRequires: Data + LabelResults: ModelTesting involves
applying MVPA model to a datasetRequires: Data + ModelResults:
Predicted LabelsTraining and Testing dataset should be independent
to avoid bias
Diagnosing DepressionSVC successfully learned patterns of
functional connectivity capable of predicting MDD from HCUncovered
differences not discovered by t-test analysisFeature selection
substantially improved the prediction accuracy of SVCMethods that
incorporate reliability performed the bestMethod requires selecting
and localizing ROIsProblematic if there is no previous research,
introduces experimenter bias/error
Predicting Intrinsic Brain ActivityMultivariate model of brain
activity
Underdetermined problem: solved using support vector regression
or other regularized regression / dimensionality reduction
methodCraddock et al. NeuroImage 2013.
Nonparametric prediction, activation, influence and
reproducibility resampling
Prediction AccuracyMeasure of the generalization ability of a
modelCan be interpreted as a measure of the information content in
the model about the region being modeled
ReproducibilityMeasures the Signal-to-Noise ratio of the
model
Strother, S. C. et al. NeuroImage 2003
Predicting Intrinsic Brain Function
Intra-individual variation
Intra-individual variation
Inter-subject prediction 480 subjects69 DZ twin pairs80 MZ twin
pairs200 Non-siblings
Train on one individual, test with anotherIntra
individualBetween siblings (MZ, DZ)Age and sex matched
non-siblings
Global prediction accuracy
Regional Differences
SVR Training
Tracking Intrinsic Connectivity Networks
Amount of Training
Predicting the Future
RT Neurofeedback of the Default Mode Network (DMN)
ICN Competition
Fox MD PNAS 2005
30
Exp. Design
Class Training Labels
Training run
Time-LabeledScans
Image Recon and SVM Classification
Image Data
Data AcquisitionStimulus Presentation
StimulusConventional FMRI
Test Data Classifier OutputTesting Run
Real-Time Tracking RSNsLaConte, et al. (2007) Hum Brain Mapp.
28: 1033-1044Stephen LaConte August 19, 2009
Stimulus seen by volunteerUpdated fMRI resultsMotion tracking
and correctionIntensity (brightness) of a single voxel, changing
during stimulus conditionsController interface for display
parameters
RT Neurofeedback of DMNTest hypothesis of DMN dysregulation in
depression, ADHD, aging, etc
Preprocessing
Online preprocessing can be performed in ~ 5 minutes, most of
which can occur in parallel with acquisition
Online DenoisingfMRI activity is confounded by intensity
modulations induced by head motion, physiological noise, scanner
drift, Implemented RT denoising in AFNI to remove contributions of
confoundsNth order polynomialGlobal meanMask average time series
(i.e. WM, CSF)Motion parameters (6 or 24 regressor models)Spatial
smoothingAdds ~ 5 ms of delay
DMN Modulation Task
Modulating the DMN
Results
Accuracy was measured from Pearsons correlation between task
paradigm and DMN activity extracted after post-processing.
Behavioral Correlates
Measures that were significantly associated with DN regulation
include (p 0.2m% Voxels with FD > 0.2m
http://preprocessed-connectomes-project.github.io/quality-assessment-protocol/
Quality Assessment Protocol (2)Implemented in pythonNormative
datasets to help learn thresholds for quality controlABIDECoRR
http://preprocessed-connectomes-project.github.io/quality-assessment-protocol/
Pledges from 25 different research institutions to share
neuroimaging data from over 3000 individuals, 500 of which suffer
from autism.
Ten new analysis software toolsets have been released from 9
institutions including MIT, the Beijing Normal University, Linkping
University (Sweden), Columbia University, and the Mind Research
Network (New Mexico).
49
High Res Anatomical
Diffusion Imaging
AcknowledgmentsCMI/NKIMichael Milham, MD, PHDZarrar ShehzadStan
Colcombe, PhDVirginia Tech Carilion Research InstituteStephen
LaConte, PhDJonathan Lisinski, MSSiemens MedicalKeith Heberlein,
PhDChris Glielmi, PhDResearch Funded in part by a NARSAD Young
Investigator Award and NIMH R01MH101555
Thank You!
Why MVPA?
Multi-voxel pattern analysis
Current state-of-the-art functional connectivity analyses are
based on mass univariate techniques that ignore the interactions
between featuresMultivariate Techniques
Null hypothesis testing evaluates p(feature|NULL), for biomarker
we are interested in p(disease state|pattern) or p(brain
state|pattern)Predictive Modeling
MVPA RegressionSimilar to classification, except label is now a
continuous variableFind a function of the features capable of
accurately producing the labels, within a specified error
Disease State PredictionTo learn a pattern of resting state
functional connectivity (FC) capable of predicting the presence or
absence of a disease state (biomarker)Apply support vector
classification (SVC) to FC maps Performed in the context of a study
of major depressive disorder (MDD)
To investigate the role of feature selection in improving
pattern prediction accuracy and interpretationIncorporate
reliability into feature selection
Subject DataData acquired on 3T Siemens TRIO Tim20 Healthy
Controls scanned with CP head coil(HC; 12F, mean age 28.9 +/-
7.2)20 depressed scanned with 8-channel matrix(MDD; 12 F, mean age
43.2 +/- 10.8)Resting State ScanEyes open and fixatingZ-Saga
sequence (20 4-mm thick axial slices, FOV = 24 cm, Matrix = 64x64,
FA/TE/TR = 90/30 ms/2.02 s)210 Volumes (7 min, 20 s)
Preprocessing / ROI extractionPreprocessing performed in
SPM5Motion correction, slice timing correction, written into MNI
space at 4 mm x 4 mm x 4 mm resolution, smoothed with 6-mm FWHM
GaussianTime course cleaning, ROI extraction performed in
AFNIRegress out residual motion, global mean, white matter, and CSF
time coursesBandpass filter (0.009 Hz < f < 0.08 Hz)ROI time
courses summarized as first eigenvariate from SVDFunctional
connectivity maps constructed by cross-correlating ROI time
coursesFisher Transform
Regions of Interest15 6-mm radius spherical ROIs chosen based on
their relevance to MDD:
Chosen from previous PET and fMRI measures of activity, not
functional connectivity
SVC of Resting State FC
Filter FS(N/TF/RF)Data Seti=0i=i+1PEii < NCalculate LOOCV
PENoWrapper FS (N/RFE/RRFE)SVCYesExclude ith observation
Observations ->Features ->
Reliability Feature Selection
T-test FilterBetween group t-test for each feature and retain
those that pass a liberal threshold (p < 0.05,
uncorrected)Univariate criterionReliability FilterEstimate
bootstrap confidence and exclude features whose 95% CI include
zero
Prediction AccuracyMethodLOOCV PEAVG # FeatAVG #
SVsRF5%1110.8TF17.5%1110.3None37.5%10530.4
Discriminating Features8 features selected by both3 features
selected by RF only due to multivariate interactions3 features
selected by TF only filtered out by reliability
TFRF
Normalized Cut (ncut) clusteringSpectral clustering methodUses
spectral (eigen) decomposition of connectivity matrix to determine
clustersInterprets connectivity matrix as a graphGraph is
partitioned into a number of subsets by removing edges from the
graphSimilarity within subsets is high Similarity between subsets
is lowNumber of clusters must be specified
ncut Cost functionNormalized cut normalizes cut cost by the
number of edges within a subsetLess sensitive to outliers than
other graph cut methods
Clustering Results