hbm parcellations
TRANSCRIPT
June 18th, 2013 OHBM morning workshop 1
Data-driven brain parcellations: A statistical perspective
Bertrand ThirionINRIA Saclay–Ile de France, PARIETAL team,
June 18th, 2013 OHBM morning workshop 2
Rationale for parcel-based data analysis K parcels rather than 105 voxels
− Multiple comparisons
− Connectivity studies
− Brain-level MVPA Local physiological parameters [Chaari et
al MICCAI 2012]
[Thirion et al HBM 2006, Compstat 2010][Craddock et al. HBM 2012]
[Varoquaux et al. Nimg 2013]
[Yeo et al. J. neurosphys. 2011]
parcel voxel cluster
June 18th, 2013 OHBM morning workshop 3
Atlases or data-driven parcellations ?
Atlases (AAL, Harvard-Oxford...) can be used to define ROIs: A priori definition and labels
Limited resolution Data-driven Parcellations:
Flexible description, better data fit
Do not fit a priori with current knowledge Lack of consistency: [Bohland et al. Plos One
2009]
June 18th, 2013 OHBM morning workshop 4
Data-driven parcellations: how ?
Any kind of data... Cyto-architecture
Sulco-gyral anatomy
Anatomical connectivity
Functional data:
− Resting-state fMRI
− Activation fMRI
− Meta-analysis / co-activation)
… many possible methods K-means, mixture models
Spectral clustering
Agglomerative clustering
Decompositions approaches:
− ICA, sparse PCA and variants
June 18th, 2013 OHBM morning workshop 5
Model selection for brain parcellations
Low K: parcels represent functional signals poorly
Large K: parcels are not reproducible
See also [Craddock et al. HBM 2012]
Model selection is an ill-posed problem- A model is not good in itself, but in view of a given objective- the data dot not conform well to models
?
June 18th, 2013 OHBM morning workshop 6
Criteria for model evaluation
(Penalized) goodness of fit− BIC criterion: penalized log-likelihood
− Cross validation: log-likelihood on left-out data (CVLL)
Reproducibility across bootstrap samples− Estimate parcellation on different subgroups and compare
co-labelling statistics (mutual information, rand index)
Voxel signal
Parcel mean signal
random subject effect
noise
June 18th, 2013 OHBM morning workshop 7
Impact of changing K on the variance
00
1.6
Variance (a.u.)
σ12
σ22
The allocation of variance into inter- and intra-subject components depends on K
− σ1
2 = within subject variance
− σ2
2 = between subject variance
June 18th, 2013 OHBM morning workshop 8
Results: goodness of fit Kopt ~ 4000 to 7000
Wards > K-means > spectral clustering
For a good summary of the activation values, use a (very) large number of parcels
BIC CV-LL
June 18th, 2013 OHBM morning workshop 9
Results: reproducibility Kopt ~ 200
Spectral clustering > Wards > K-means
To reproduce well the parcels, use ~ 200 parcels
Accuracy
Reproducibility
June 18th, 2013 OHBM morning workshop 10
Hints from simulations
Poor between subject registration might artificially inflate the number of parcels required to fit the signal
Functional registration should improve the estimation [Sabuncu et al. Cerb. Cortex 2009, Robinson et al. IPMI 2013 ]
Smoothing also inflates the number of parcels
June 18th, 2013 OHBM morning workshop 11
Discussion
Current atlases are too coarse to yield reliable averages of fMRI data
Goodness of fit is different from stability / reproducibility [Strother et al. 2002]
Wards' methods better suited than alternatives
June 18th, 2013 OHBM morning workshop 12
What about resting state Consider linear decompositions and clustering
The signal cannot be easily modeled probabilistically
Cross-validation of the R2 of resting-state signals, AMI
Smaller number of regions (~80) [Abraham et al MICCAI 2013]
Accuracy Reproducibility
June 18th, 2013 OHBM morning workshop 13
Resting-state parcellations
[Abraham et al MICCAI 2013]
June 18th, 2013 OHBM morning workshop 14
Conclusion
Usefulness of brain parcellations
− A good model depends on the context
− Reproducibility and accuracy yields different responses Need (more) multi-modal data to properly define regions
Winners:
− Ward's clustering (large K)
− Dictionary learning (small K) Might be worth combining results from different parcellations
[Varoquaux et al. ICML 2012, da Mota et al. MICCAI 2013, Poster #1275]
June 18th, 2013 OHBM morning workshop 15
Acknowledgements
Gaël Varoquaux, Alexandre Abraham, Alan Tucholka, Benoit da Mota, Virgile Fritsch, Vincent Michel
JB Poline, Guillaume Flandin, Philippe Pinel