This document summarizes research predicting autism from resting-state fMRI data using a connectome classification pipeline. The pipeline involves defining regions of interest, extracting time series, computing functional connectivity matrices, and using supervised learning. The authors explore different choices for each step and find that learning regions with MSDL, using tangent-space embedding for connectivity, and standard SVM learning work best across datasets with different heterogeneity levels. The findings suggest connectome structure is less important for prediction than choice of regions and preprocessing.
2. Predictive biomarkers from rest
Suitable for diminished patients
Probing intrinsic brain structure
Predicting pathologies as a validation proxy
Multi-site large autism dataset: ABIDE
Autism Spectrum Disorder
[Di Martino... 2014]
⇒ Patient/Control classification
16 sites
∼ 1200 subjects (900 after QC)
G Varoquaux 2
3. Experiments: autism prediction
Seen sites: predicting to new subjects
Training set Testing set
Unseen sites: predicting to new sites
Training set Testing set
Training set Test set
G Varoquaux 3
4. A connectome classification pipeline
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
G Varoquaux 4
5. A connectome classification pipeline
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
Prediction accuracy (%)
Seen sites 67±3
Unseen sites 67±5
What is important to predict?
G Varoquaux 4
6. A connectome classification pipeline
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
Prediction accuracy (%)
Seen sites 67±3
Unseen sites 67±5
What is important to predict?
G Varoquaux 4
8. 1 ROI definition: Learning functional regions
Harvard Oxford
Anatomical atlases do not resolve functional
structures
Where is the default mode network?
G Varoquaux 6
9. 1 ROI definition: Clustering approaches
K-Means
Related to [Yeo... 2011]
Normalized cuts
[Craddock... 2012]
Ward clustering
[Thirion... 2014]
G Varoquaux 7
10. 1 ROI definition: Linear decomposition approaches
Model: Observing linear mixtures of networks at rest
Time courses
G Varoquaux 8
11. 1 ROI definition: Linear decomposition approaches
Model: Observing linear mixtures of networks at rest
Time courses
Language
G Varoquaux 8
12. 1 ROI definition: Linear decomposition approaches
Model: Observing linear mixtures of networks at rest
Time courses
Audio
G Varoquaux 8
13. 1 ROI definition: Linear decomposition approaches
Model: Observing linear mixtures of networks at rest
Time courses
Visual
G Varoquaux 8
14. 1 ROI definition: Linear decomposition approaches
Model: Observing linear mixtures of networks at rest
Time courses
Dorsal Att.
G Varoquaux 8
15. 1 ROI definition: Linear decomposition approaches
Model: Observing linear mixtures of networks at rest
Time courses
Motor
G Varoquaux 8
16. 1 ROI definition: Linear decomposition approaches
Model: Observing linear mixtures of networks at rest
Time courses
Salience
G Varoquaux 8
17. 1 ROI definition: Linear decomposition approaches
Model: Observing linear mixtures of networks at rest
Time courses
Observing a mixture
How to unmix networks?
ICA: minimize mutual information across S
Sparse decompositions: sparse penalty on maps
G Varoquaux 8
18. 1 ROI definition: MSDL linear decomposition
[Varoquaux... 2011, Abraham... 2013]
MSDL: multi-subject dictionary learning
Sparse
Multi-subject
Subject maps + Population-level
atlas
Spatial penalty
(total variation)
G Varoquaux 9
19. 1 ROI definition: MSDL linear decomposition
[Varoquaux... 2011, Abraham... 2013]
MSDL: multi-subject dictionary learning
Sparse
Multi-subject
Subject maps + Population-level
atlas
Spatial penalty
(total variation)
L R
y=-23 x=0
L R
z=18
L R
y=-78 x=2
L R
z=4
Ventricular system Visual Cortex
L R
y=-47 x=51
L R
z=11
L R
y=-17 x=-47
L R
z=8
TPJ Auditory Network
G Varoquaux 9
27. 3 Functional-connectivity matrix: impact of choice
Time series
2
RS-fMRI
41
Diagnosis
ROIs Functional
connectivity
3
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
G Varoquaux 15
28. 4 Supervised learning method
Functional
connectivity
Time series
3
4
Diagnosis
2
RS-fMRI
1 ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
G Varoquaux 16
29. 4 Supervised learning method
Functional
connectivity
Time series
3
4
Diagnosis
2
RS-fMRI
1 ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
Standard supervised learning
Ridge classifier
SVC 2 penalized
SVC 1 penalized
G Varoquaux 16
30. 4 Supervised learning method: impact of choice
Functional
connectivity
Time series
3
4
Diagnosis
2
RS-fMRI
1 ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
G Varoquaux 17
31. Importance of pipeline steps
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
G Varoquaux 18
32. Importance of pipeline steps
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
G Varoquaux 18
33. How general are these conclusions?
Different subsets:
Set Card. Sites Selection criteria
A 871 17 All subjects after QC
B 736 11 Remove 6 smallest sites
C 639 14 A without left handed and women
D 420 14 C between 9 and 18 years old
E 226 3 D from 3 biggest sites
Exploring variability
From most heterogeneous
to most homogeneous
G Varoquaux 19
34. Across different sets of subjects
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
MSDL outperform alternatives
Intra-site
Inter-site
G Varoquaux 20
35. Across different sets of subjects
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
Tangent space outperform alternatives
Intra-site
Inter-site
G Varoquaux 20
36. Across different sets of subjects
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
1 ROI definition
2 Time-series extraction
3 Connectivity matrices
4 Supervised learning
2 classifiers outperform alternatives
Intra-site
Inter-site
G Varoquaux 20
40. @GaelVaroquaux
Intersite biomarkers from rest [Abraham submitted]
Prediction of autism
Carries out across sites
Recipe for a good pipeline
Choice of regions critical (MSDL to learn them)
Tangent-space embedding
Standard SVM
Connectome structure not central for prediction
41. @GaelVaroquaux
Intersite biomarkers from rest [Abraham submitted]
Prediction of autism
Carries out across sites
Recipe for a good pipeline
Choice of regions critical (MSDL to learn them)
Tangent-space embedding
Standard SVM
Definition of regions MSDL
Linear decomposition + sparsity
+ total-variation
ni
42. References I
A. Abraham, E. Dohmatob, B. Thirion, D. Samaras, and
G. Varoquaux. Extracting brain regions from rest fMRI with
total-variation constrained dictionary learning. In MICCAI, page
607. 2013.
R. C. Craddock, G. A. James, P. E. Holtzheimer, X. P. Hu, and
H. S. Mayberg. A whole brain fMRI atlas generated via spatially
constrained spectral clustering. Human brain mapping, 33(8):
1914–1928, 2012.
A. Di Martino, C.-G. Yan, Q. Li, E. Denio, F. X. Castellanos,
K. Alaerts, J. S. Anderson, M. Assaf, S. Y. Bookheimer,
M. Dapretto, ... The autism brain imaging data exchange:
towards a large-scale evaluation of the intrinsic brain
architecture in autism. Molecular psychiatry, 19:659, 2014.
B. Thirion, G. Varoquaux, E. Dohmatob, and J. Poline. Which
fMRI clustering gives good brain parcellations? Name: Frontiers
in Neuroscience, 8:167, 2014.
43. References II
G. Varoquaux, F. Baronnet, A. Kleinschmidt, P. Fillard, and
B. Thirion. Detection of brain functional-connectivity difference
in post-stroke patients using group-level covariance modeling. In
MICCAI, pages 200–208. 2010.
G. Varoquaux, A. Gramfort, F. Pedregosa, V. Michel, and
B. Thirion. Multi-subject dictionary learning to segment an atlas
of brain spontaneous activity. In Inf Proc Med Imag, pages
562–573, 2011.
B. Yeo, F. Krienen, J. Sepulcre, M. Sabuncu, ... The organization
of the human cerebral cortex estimated by intrinsic functional
connectivity. J Neurophysio, 106:1125, 2011.