Multi-subject Bayesian joint detection & estimation in fMRI Philippe Ciuciu

advertisement
1/31
Multi-subject Bayesian joint
detection & estimation in fMRI
Philippe Ciuciu
joint work with Solveig Badillo
CEA – NeuroSpin
INRIA – Parietal Team
NeuroStats 2014 WS, Sep. 3-5, 2014
University of Warwick
2/31
Brain Activity Measurement in fMRI
fMRI: BOLD effect
Blood oxygenation level dependent signal
[Ogawa et al., PNAS 1990]
3/31
Motivations
→ HRF
shape estimation at the group-level:
●
●
●
Group comparisons (patients vs controls, young vs elderly subjects)
Effect of treatment (Placebo vs Drug) in clinical trials
Compare hemodynamic variability sources (region, condition, ...)
hS 1
hS 2
hS3
Outline
●
Current approaches
●
Bayesian joint detection & estimation (JDE)
formalism
●
Multi-subject extension
●
Results
●
Conclusions
4/31
5/31
Group-level HRF estimation
Classical approach:
1/ Perform subject-by-subject estimation
2/ Average over subjects
→ not robust to outliers
6/31
Standard brain mapping analysis
General Linear Model:
[Friston et al., NeuroImage 1995]
Effects of interest
7/31
HRF recovery within the GLM
General Linear Model (SPM)
[Frackowiak et al., Hum. Brain Func. 1997]
●
Finite Impulse Response (FIR)
HRF variability :
➢
Within-subject, between-subjects,
between-groups (children, patients, …)
[Miezin et al., NeuroImage 2000;
D'Esposito et al, NeuroImage 2003;
Handwerker et al, NeuroImage 2004]
[Henson et al., Hum. Brain Func. - Chap. 10]
8/31
HRF
withinin
the
GLM
Tworecovery
main issues
fMRI
●
Limits of the FIR model:
●
Synchronous paradigms
●
Putative loss of sensitivity
●
Lack of robustness
[Henson et al., Hum. Brain Func. - Chap. 10]
●
Interplay between detection and estimation
Joint detection and estimation (JDE)
[Makni et al., IEEE TSP 2005, NeuroImage 2008]
9/31
Group-level HRF estimation
Multi-subject inference:
●
Semi-parametric univariate estimation framework
[Zhang et al., NeuroImage 2013]
●
●
Semi-parametric joint Detection-Estimation
univariate framework [Degras & Lindquist, NeuroImage 2014]
Non-parametric Bayesian JDE multivariate framework
[Badillo et al, IEEE PRNI 2014]
Outline
●
Current approaches
●
Bayesian joint detection & estimation (JDE) formalism
●
Multi-subject extension
●
Results
●
Conclusions
10/31
11/31
Forward Modeling of fMRI Time Series
[Makni et al., NeuroImage 2008]
[Vincent et al., IEEE TMI 2010]
12/31
Whole Brain Analysis
13/31
●
●
●
Hybrid MH-Gibbs sampling
Posterior mean estimates
Faster VEM alternative
[Vincent
[Vincent et
et al.,
al., IEEE
IEEE TMI
TMI 2010]
2010]
Risser
et
al.,
JSPS
2011]
Risser et al., JSPS 2011]
[Chaari et al., IEEE TMI 2013]
14/31
Within-subject Results
Parietal and frontal HRFs
T-score
JDE
MCMC & VEM
-
7.5
5
2.5
0
- -2.5
GLM
- -5
%Δsignal BOLD
Computation – Sentence contrast
- -7.5
Time (s.)
[Chaari et al, IEEE TMI 2013; Badillo et al, NeuroImage 2013]
Outline
●
Current approaches
●
Bayesian joint detection & estimation (JDE) formalism
●
Multi-subject extension
●
Results
●
Conclusions
15/31
16/31
Multi-Subject JDE Forward Model
data
drift
noise
voxel j
conditions m=1:M
subject s=1:S
subject-specific
Group
parcellation
17/31
MS-JDE Supplementary Prior
data
drift
subject-specific
HRF Priors:


noise
18/31
MS-JDE Supplementary Prior
data
drift
subject-specific
HRF Priors:

Potential subject-specific shift of the peak
can be generalized to other population
distribution (eg, mixtures, Student)
noise
19/31
MS-JDE Directed Acyclic Graph
MCMC implementation: see
[Badillo et al, IEEE PRNI 2014]
for details
Outline
●
Current approaches
●
Bayesian joint detection & estimation (JDE) formalism
●
Multi-subject extension
●
Results
●
Conclusions
20/31
21/31
Simulated Data
Simulation of 4D data:
→ Fast event-related paradigm
TR=1s; N=300 scans, SOA~3.6s, 1run
→ Simulation of activated areas
q=0
✔
q=1
✔
400 time points
Response levels:
q=0
→ Group-level HRF
→ Subject-level HRF
q=1
✔
Drift:
✔
Noise:
22/31
Validation on simulations
subject-specific HRF
GT: Ground Truth
SS-JDE: Single-subject JDE
MS-JDE: Multi-subject JDE
23/31
Validation on simulations
subject-specific HRF
subject-specific HRF
subject-specific HRF
GT: Ground Truth
SS-JDE: Single-subject JDE
MS-JDE: Multi-subject JDE
24/31
Validation on simulations
subject-specific HRF
group-level HRF
subject-specific HRF
GT: Ground Truth
SS-JDE: Single-subject JDE
MS-JDE: Multi-subject JDE
subject-specific HRF
25/31
Localizer Dataset
­ Fast event­related paradigm, 1 run, 400 scans, TR=1s, mband­EPI sequence (CMRR, Mineapolis), mb=3, no­PI, 3x3x3 mm3, SOA~3.6 s
­ 10 conditions, visual or auditory modality
26/31
Results on Elderly Subjects
%Δsignal BOLD
group-level HRF
shape
15 elderly adults
(65 to 75 years)
Time (s.)
elderly
SS­JDE: TTP=6 s
FWHM=7.2 s
MS­JDE: TTP=5.8 s
FWHM=7.6 s
27/31
Results on Elderly Subjects
%Δsignal BOLD
%Δsignal BOLD
group-level HRF
shape
15 elderly adults
(65 to 75 years)
Time (s.)
elderly
SS­JDE: TTP=6 s
FWHM=7.2 s
MS­JDE: TTP=5.8 s
FWHM=7.6 s
Time (s.)
elderly
SS­JDE: TTP=6 s
FWHM=6 s
MS­JDE: TTP=7.5 s
FWHM=8.5 s
28/31
Results on Elderly Subjects
%Δsignal BOLD
%Δsignal BOLD
group-level HRF
shape
15 elderly adults
(65 to 75 years)
Time (s.)
Time (s.)
MS-JDE model captures inter-subject variability
29/31
Group comparison
group-level HRF
shapes
15 elderly adults vs
15 young adults
MS-JDE Young subjs.
MS-jDE Elderly subjs.
%Δsignal BOLD
%Δsignal BOLD
MS-JDE Young subjs.
MS-jDE Elderly subjs.
Time (s.)
Young adults: TTP=5.5 s
FWHM=6 s
Elderly adults: TTP=5.8 s
FWHM=7.6 s
Time (s.)
Young adults: TTP=6.7 s
FWHM=7.6 s
Elderly adults: TTP=7.5 s
FWHM=8.5 s
30/31
The PyHRF software
http://pyhrf.org
Nipy
Nibabel
...
soma-workflow
XML parametrization
Visualization
31/31
Summary
Multi-subject HRF estimation can be helpful for group
➢
comparison
➢
Joint detection & estimation is the right framework!
✗
✗
✗
➢
Neither necessarily Bayesian nor non-parametric HRF
But multivariate analysis of evoked activity is crucial
Main issue: multi-subject parcellation
More complex random-effect models can be envisaged
Computational bottleneck: parallelize all steps that you
can!
➢
➢
Perspectives :
✗
✗
✗
Comparison with B-spline HRF decomposition
Multi-subject JPDE
Assess the relevance of hemodynamics characteristics as
potential biomarkers in clinical trials (phMRI)
Download
Random flashcards
Arab people

15 Cards

Radioactivity

30 Cards

Nomads

17 Cards

Marketing

46 Cards

Ethnology

14 Cards

Create flashcards