E-Poster

A Bayesian GLM for Cortical Surface fMRI Task Analysis
Amanda F.
1
Mejia ,
Finn
4
Lindgren
and Martin A.
methods
corticalsurfacefMRI
L
results
1
β€’ Reduceddimensionality
β€’ Improvedalignmentofcorticalareas
β€’ Moremeaningfulsmoothing&
spatialmodeling
2
AssumeSPDE
prioroneach𝜷
3
UseINLA
toestimate
posteriors
Identifyareasof
activationusing
jointPPM
Jointgroup-level
model
R
1. Stochastic partial differential equation prior (Lindgren et al. 2011)
fMRItimeseries
atlocation𝑣
Taskdesign Activationat Nuisance β€œActivation”due
signals
tonuisanceat
location𝑣
matrix
location𝑣
𝑦: = 𝑋𝛽: + 𝑍𝑏: + πœ–:
(𝑇×1)
(𝑇×𝐾) 𝐾×1 (𝑇×𝐽) 𝐽×1 (𝑇×1)
𝑇 – lengthoffMRItimeseries
𝑉 – numberoflocations
𝐾 – numberoftasks+baseline
𝐽 – numberofnuisancesignals
fMRItime
seriesatall
locations
- Finite approximation to a continuous Matérn Gaussian random field
- Markov property due to sparse inverse covariance structure
𝑓 𝒖
𝑓 𝒖 β‰ˆ P πœ“R 𝒖 𝑀R
- Constructed on a triangular mesh
R
(format of cs-fMRI data)
BayesianGLM
Taskdesign Activationat Nuisance β€œActivation”due
signals tonuisanceatall
alllocations
matrix
locations
π‘Œ = 𝑋𝛽 + 𝑍𝐡 + 𝐸
(𝑇×𝑉)
(𝑇×𝐾) 𝐾×𝑉 (𝑇×𝐽) 𝐽×𝑉 (𝑇×𝑉)
2. Integrated nested Laplace approximation (Rue et al. 2009)
𝛽 ∼ πœ‹(𝛽|πœƒ)
Spatialprocess
prior(e.g.GMRF)
- Accurately estimates posterior densities using Laplace approximation
- Much faster than MCMC for Bayesian computation
numerical
- Implemented in R (www.r-inla.org)
Hyperparameters
Pitfallsofβ€œmassiveunivariate” AdvantagesofBayesianGLM:
β€’ Smoothercoefficientestimates
classicalGLMapproach:
πœ‹ π›½π‘˜ 𝑦 = ∫ πœ‹ π›½π‘˜ πœƒ, 𝑦 πœ‹ πœƒ 𝑦 π‘‘πœƒ
β€’ Degreeofsmoothnessdetermined
β€’ Failstoaccountforspatial
separatelyforeachactivationfield
dependenceinactivation
β€’ Possibletoavoidmultiple
β€’ SmoothingincreasesSNR,but
comparisonscorrection,sinceall
shapeanddegreead-hocand
locationsmodeledjointly
commonacrosstasks
β€’ Multiplecomparisoncorrection β€’ Morepowertodetectactivation
essentialforfalsepositive
ChallengesofBayesianGLM:
control
1. Assumeanappropriateprioron𝛽s
- Complexspatialandtemporal
dependenciesmakepropercorrection
challenging
- Smoothingincreasesdependence
betweentests
- Populartechniqueshavebeenfound
tolackpowerorhavehighfalse
positiverates
- Excursion set is the largest set of locations 𝐸𝛾,𝛼 such that every location
exceeds activation level 𝛾 with joint probability 1 βˆ’ 𝛼
- Excursion function is the largest value 1 βˆ’ 𝛼 for which each location
would be included in 𝐸𝛾,𝛼 - Estimate excursion function using INLA techniques, then threshold to
identify regions of activation at a given significance level 𝛼
2. PerformBayesiancomputation
Group-level model
- MCMCslow,maymixpoorly
- VBfastbutunderestimatesvariance
3. Thresholdposteriorprobabilitymap
(PPM)todetectareasofactivation
- Define group-level activations 𝛽𝐺 as linear combination (e.g., mean) of
subject-level 𝛽𝑖 ’s: 𝛽𝐺 = 𝐴𝛽, 𝛽 = 𝛽1β€² , … , 𝛽𝑛′ β€²
- Estimate posterior of 𝛽𝐺 using importance sampling
useINLAGaussianapprox.
β€’
β€’
fromsubject-levelmodels
πœ‹(𝛽𝐺 𝑦 = ∫ πœ‹(𝛽𝐺 𝑦, πœƒ πœ‹ πœƒ 𝑦 π‘‘πœƒ
4. Modeldatafrommultiplesubjects?
Use original subject-level surfaces for more accurate
distance-based spatial dependence modeling
Relax assumption of stationarity in SPDE prior
Use empirical spatial dependence structure for each
task to improve computational efficiency and
incorporate long-range dependencies
\
3. Identify areas of activation with joint PPM (Bolin & Lindgren 2015)
- Shouldcapturespatialdependence
structurewhilehavingsparseprecision
- cs-fMRIdatahasmeshformat,butmany
priorsassumelatticestructure
future work
approximation
πœ‹ 𝛽, πœƒ, 𝑦
πœ‹U πœƒ 𝑦 ∝
X
πœ‹UW 𝛽 πœƒ, 𝑦 YZYβˆ—
- MarginalPPMconvenientbutreintroduces
multiplecomparisonsissue
β€’
simulation study
β€’ Generated a time series of length 200 for a single slice consisting of
baseline (GM probability map) + two activation profiles + AR(1) noise
β€’ Fit classical GLM (after smoothing & prewhitening) and Bayesian GLM with
AR(1) errors
β€’ Classical GLM: corrected for multiple comparisons by
- FDR correction using Benjamini-Yekutieli procedure (π‘ž = 0.01)
- FWER correction using a permutation test (𝛼 = 0.01)
β€’ Bayesian GLM: marginal and joint PPMs used to ID activations (𝛼 = 0.01)
Estimateasamixtureof
Gaussiansforeach
samplefromπ‘ž πœƒ 𝑦
βˆπœ‹ πœƒ
]^_
` πœ‹ πœƒ 𝑦a
a
∝Gaussiandensityπ‘ž πœƒ 𝑦
HCP motor task study
β€’ 20 randomly sampled subjects from HCP 500 performed 5 motor tasks (12
sec) following visual cue (3 sec) over 3.5 min (284 volumes)
β€’ Data minimally preprocessed according to HCP fMRISurface pipeline, plus:
- Classical GLM: surface smoothing (6mm FWHM) using Connectome
Workbench
- Bayesian GLM: resampling (32K to 6K) using Connectome Workbench
β€’ Prewhitened data using an AR(6) model with spatially-varying coefficients
β€’ Included 14 nuisance covariates in models to account for motion and drift
β€’ To identify areas of activation, used FDR and FWER correction as in
simulation for classical GLM, joint PPM approach for Bayesian GLM at 0.01
Subjectestimates
ClassicalGLM
ClassicalGLM
5
Lindquist
Groupestimates
Areasofactivation(𝛼, π‘ž = 0.01)
FWER
FDR
JointPPM(𝛼 = 0.01)
BayesianGLM
β€’ Veryhighdimensional
β€’ Difficulttoaligncorticalareas
β€’ Complexspatialdependencies
1 Indiana
David
3
Bolin ,
University, Bloomington, Indiana, USA
2 Baruch College, The City University of New York, New York, USA
3 Gothenburg University, Gothenburg, Sweden
4 The University of Edinburgh, Edinburgh, UK
5 Johns Hopkins University, Baltimore, Maryland, USA
background
volumetricfMRI
Yu Ryan
2
Yue ,
𝛾=0
𝛾 = 1%baseline
references
β€’
β€’
β€’
β€’
β€’
Mejia AF, Yue Y, Bolin D, Lindgren F & Lindquist MA (2017+). A Bayesian GLM approach to cortical surface fMRI data analysis.
Lindgren F, Rue H and Lindström J (2011). An explicit link between Gaussian fields and GMRFs: the SPDE approach (with discussion). JRSSB.
Whittle, P. (1954). On stationary processes in the plane. Biometrika.
Rue H, Martino S and Chopin N (2009). Approximate Bayesian inference for latent Gaussian models using integrated nested Laplace
approximations (INLA) (with discussion). JRSSB.
Bolin D and Lindgren F (2015). Excursion and contour uncertainty regions for latent Gaussian models. JRSSB.