the PyHRF package P. Ciuciu1,2 and T. Vincent1,2 Methods meeting at Neurospin 1: CEA/NeuroSpin/LNAO

Transcription

1 Joint detection-estimation of brain activity from fmri time series: the PyHRF package Methods meeting at Neurospin P. Ciuciu1,2 and T. Vincent1,2 1: CEA/NeuroSpin/LNAO 2: IFR49

2 Outline 2/61 I. Introduction II. A novel framework for intra-subject analysis in fmri III. Implementation : the PyHRF software IV. Conclusions and perspectives

3 3/61 Classical fmri analysis 1. Detect and localise brain activations Ex: In SPM [Friston et al, 1994], the BOLD response is modelled with: or Compute statistical activation Maps 2. Estimate the dynamics of activation [Goutte et al, IEEE TMI 2000; Marrelec et al, HBM 2003] HRF estimates Probe brain dynamics Time s

4 4/61 Classical fmri analysis GLM limitations: A single HRF shape is not physiologically appropriate Studies report HRF variability: within subject (between regions, sessions, conditions, trials) [Miezin et al., NIM 2000; Ciuciu et al., IEEE TMI 2003] [Neumann et al, NIM 2003, 2006; Smith et al, NIM 2005] between subjects [Handwerker et al., NIM 2004; Aguirre et al., NIM 1998] between groups (infants, patients,...)... [D'Esposito et al, NIM 1998, 2003; Richter and Richter, NIM 2003]

5 5/61 Motivations Two main issues in fmri Detection of brain activation and estimation of brain dynamics are addressed separately Any detection method supposes a given HRF shape Any estimation algorithm provides relevant results in activated voxels or regions only Address these two problems simultaneously in a joint detection-estimation (JDE) framework [Makni et al., IEEE SP 2005, ISBI'06, ICASSP'06, NIM, 2008; Vincent et al, ICASSP'07, EMBC'07, ISBI'08; Ciuciu et al, ISBI'08 ; Risser et al, sub. to NIPS'08; Vincent et al, sub. to IEEE TMI]

6 Outline 6/61 I. Introduction II. A novel framework for intra-subject analysis in fmri III. Implementation : the pyhrf software IV. Conclusions and perspectives

7 Forward BOLD model 7/61 [Makni, Ciuciu et al., IEEE SP 2005; Makni et al, Ciuciu, NeuroImage, 2008] Region definition: voxels Region of interest fmri time series in voxels of the ROI Experimental paradigm: Arrival times of M stimuli (onsets)

8 8/61 Forward BOLD model Temporal hypotheses: for standard ISIs Hyp. 1: linearity Hyp 2: stationarity + Convolution kernel Hyp 3: additivity Time in s TR 2TR3TR... Stimulus-varying NRLs:

9 Forward BOLD model 9/61 Spatial hypotheses: functionally homogeneous ROI Single HRF shape Voxel-dependent magnitudes of the BOLD response Neural Response Levels

10 Forward BOLD model 10/61 Spatial hypotheses: functionally homogeneous ROI Single HRF shape [Makni, Ciuciu et al., IEEE SP 2005; Makni et al, Ciuciu, NeuroImage, 2008] Voxel-dependent magnitudes of the BOLD response

11 11/61 Forward BOLD model Unknown parameters NRL in and for condition = = BOLD signal measured in voxel Arrival time of stimulus Noise statistics Drift coefficients in voxel HRF Orthonormal basis for low frequency drift modelling Known parameters

12 Bayes rule 12/61 likelihood How the data are generated from the parameters? Forward modeling

13 13/61 Likelihood definition Main hypothesis: noise decorrelated in space fmri time series are statistically independent in space: Temporal noise model: either white or serially correlated AR(1) [Makni et al, Ciuciu, NeuroImage 2008]

14 14/61 Bayes rule Prior What do we know about the parameters before the data are acquired? Prior modeling

15 HRF prior modeling 15/61 Information about the HRF shape nonparametric approaches Averaging: [Buckner et al, 1996] Selective averaging: [Dale et al, HBM 1997] FIR (Finite Impulse Response) Regularized FIR (smoothing prior): [Marrelec, Ciuciu et al, IPMI 03 Ciuciu et al., IEEE TMI 2003]

16 16/61 NRL prior modeling Spatial information [Vincent, Ciuciu et al, ICASSP'07] Independence between conditions: Spatial mixture model (SMM) for each Discrete Markov random Field (eg Potts/Ising): Partition function:

17 17/61 Prior model Two-class Gaussian mixture [Vincent et al., ICASSP' 07] Non-activating voxels: Activating voxels: Unknown hyper-parameters: Influence of

18 18/61 Bayes rule What do we know about the HRF, the NRLs and the hyper-parameters given the data? Keystone of learning scheme: simulating realizations of using Gibbs sampler

19 19/61 Real data sets Localizer fmri experiment: Experimental conditions under study: auditory and visual stimuli Event-related paradigm : - Short stimuli duration - Inter-stimulus interval : ~3s to 10s - Randomised sequence 125 scans with TR = 2.4s, scanning at 3T ROIs located in the primary auditory cortices

20 20/61 SPM vs. JDE Auditory Visual contrast: % Δ BOLD signal JDE SPM % Δ BOLD signal Time in s. Time in s. Bilateral activation detected along the gray matter from raw data sets (spatially unsmoothed)

21 21/61 Whole brain analysis First step: Segmentation of the gray/white matter interface from the T1 MRI Second step : brain parcellation based upon functional similarities and spatial connectivty [Flandin et al, ISBI'02; Thirion et al, HBM 2006] [Makni et al, Ciuciu, NIM 2008]

22 Sources of variability Subject, region 22/61 [Birn et al, NIM 2001; Marrelec et al, HBM 2003; Menon et al NIM 2003; Neumann et al NIM 2003; Handwerker et al, NIM 2004] [Makni et al, Ciuciu, NIM 2008]

23 Whole brain: supervised SMM =0.1 =0.3 23/61 =0.7

24 Outline 24/61 I. Introduction II. A novel framework for intra-subject analysis in fmri III. Implementation : the PyHRF software IV. Conclusions and perspectives

25 PyHRF package 25/61 Outline General framework Implementation and requirements Setup and use cases Viewer overview

26 fmri analysis with PyHRF 26/61 t t t PyHRF Parcellation = n-ary mask (0 label = background) from SPM beta maps from 4D fmri series

27 fmri analysis with PyHRF 27/61 t t t HRF PyHRF Parcellation = n-ary mask (0 label = background) from SPM beta maps from 4D fmri series contrasts

28 Package description 28/61 Mainly command line Setup with XML files Interactive dedicated viewer Python and C Time issues: parcellation : several minutes JDE : 1 to 2 hours (can be parallelized) Available at: $P4/neursopy-.../ $python setup.py install

29 29/61 Dependencies gsl scipy numpy fff PyHRF clustering viewer PyQt JDE PyXML pynifti SIP Qt nifticlib

30 30/61 Dependencies gsl scipy numpy fff PyHRF clustering viewer PyQt JDE PyXML pynifti SIP Qt nifticlib

31 Dependencies (versions) 31/61 fff (snapshot SVN may 2008) GSL 1.9 numpy 1.0 scipy PyXML nifticlib 1.0 pynifti PyQt Qt SIP 4.7

32 Use cases 32/61 Parcellation JDE Viewer

33 Parcellation 33/61 $pyhrf_clus_buildcfg clustering.xml $pyhrf_clus_run parcellation.nii

34 Use cases 34/61 Parcellation JDE Viewer

35 35/61 Joint Detection-Estimation $pyhrf_jde_buildcfg detectestim.xml $pyhrf_jde_estim -v1 session0_pm_hrf.nii session0_pmnrl_condition_audio.nii session0_pmnrl_condition_audio.nii...

36 Joint Detection-Estimation 36/61 $pyhrf_jde_buildcfg detectestim.xml Data definition Model parameters $pyhrf_jde_estim -v1 session0_pm_hrf.nii session0_pmnrl_condition_audio.nii session0_pmnrl_condition_audio.nii...

37 Joint Detection-Estimation 37/61 detectestim.xml Data definition - seconds - separator : space parcellation with positions excluded during analysis (output file) Input parcellation BOLD data - seconds - separator : space - if empty : spike stimuli

38 Joint Detection-Estimation 38/61 detectestim.xml Model definition (1) Drift base type ('polynomial' or 'cosine') Drift parameter: polynom order or cut-off frequency

39 Joint Detection-Estimation 39/61 detectestim.xml Définition du modèle (2) Amount of spatial correlation (between 0.3 et 0.7) Contrast definition Separator: ; If no contrast: ; Boolean flags for outputs Constraint of HRF=0 at first and last time position (boolean) Boolean flag for the HRF estimation. If 0 : use of a fixed canonical HRF

40 40/61 Joint Detection-Estimation Multi-session aspect $pyhrf_jde_buildcfg -s homoscedastic $pyhrf_jde_buildcfg -s independent Two different xml setup files

41 41/61 Joint Detection-Estimation $pyhrf_jde_buildcfg detectestim.xml $pyhrf_jde_estim -v1 session0_pm_hrf.nii session0_pmnrl_condition_audio.nii session0_pmnrl_condition_audio.nii...

42 Use cases 42/61 Parcellation JDE Viewer

43 Viewer 43/61 $pyhrf_view session0*

44 44/61 Viewer $pyhrf_view session0* Data browser (slice definition) current slice view

45 45/61 Viewer Data browser (slice definition) current slice view Slice dimension

46 46/61 Viewer Data browser (slice definition) current slice view Crop to mask

47 47/61 Viewer Data browser (slice definition) current slice view Slice axes selection

48 48/61 Viewer Data browser (slice definition) current slice view Mode 2D: image (matrix) Mode 1D: curves

49 PyHRF package in factory 49/61 Methodological developments Unsupervised spatial regularization Habituation modeling Model comparison Software developments Integration to the fmri Brainvisa toolbox User friendly parallel system Scriptable API Port of the matlab HRF toolbox: L. Risser's work.

50 Conclusions 50/61 The joint detection-estimation framework: directly accounts for different sources of variability provides both region-based HRF time courses and contrast maps No spatial filtering/smoothing of fmri datasets depends on an input parcellation: [Vincent et al, ISBI'08] gives improved RFX maps in comparison with SPM First release of Pyhrf package (v 1.0) available either in the BrainVisa software or as a standalone toolbox.

51 Extract temporal information 51/61 Multiple conditions Primary auditory cortex Time in s Planum temporale Time in s [Ciuciu et al, IEEE TMI 2003]

52 Sensitivity analysis 52/61 [Vincent, Ciuciu, ISBI 2008] Random parcellation JDE Single JDE run GMM parcellation JDE

53 Sensitivity analysis 53/61 [Vincent, Ciuciu, ISBI 2008] Random parcellations JDE JDE JDE JDE... Single JDE run GMM parcellation JDE

54 Whole brain: supervised SMM =0.1 =0.3 54/61 =0.7

55 55/61 Supervised vs unsupervised SMM unsup. =0.1 =0.7

56 Adaptive spatial regularization 56/61 RW Metropolis-Hastings step for sampling Preliminary estimation of the partition function Fondamental identity: Practical implementation: Tabulate over a fine grid For Generate of (SW scheme) Compute [Meng and Rubin BIOM 1998, Higdon et al, IEEE TMI 1997]

57 Why modeling the HRF? 57/61 Motivations Extract temporal information Account for variability sources Gain in sensitivity of detection [Marrelec et al, HBM 2003] Compare BOLD fmri sequences (EPI vs EVI) [Rabrait et al, JMRI 2008] Coherent approach for EEG/fMRI fusion [Daunizeau et al, IEEE SP 2005, NIM 2006/07]

58 58/61 Adaptive spatial regularization Remarks: Extrapolation method for ROI of variable size and shape Parcel-dependent regularization factor [Risser, Vincent and Ciuciu, sub. to NIPS 08]

59 59/61 Supervised vs unsupervised SMM unsup. =0.1 =0.7

60 60/61 fmri analysis with PyHRF t t t PyHRF SPM = Parcellation

61 61/61 fmri analysis with PyHRF t t t HRF PyHRF SPM = contrasts Parcellation

62 62/61 Implemented model Supervised spatial mixture modeling for neural responses White noise Polynomial/cosine low frequency drift (no output, integrated out of the problem) Gausian HRF which covariance is constraint by its second derivative (smooth constraint) Homoscedastic multisession modeling