A tutorial on using machine-learning for functional-connectomes, for instance on resting-state fMRI. This is typically useful for population imaging: comparing traits or conditions across subjects.

1. Machine learning for functional connectomes
Gaël Varoquaux

2. Machine learning for functional connectomes
Gaël Varoquaux
Outline:
1 Intuitions on machine learning
2 Machine learning on rest fMRI
Pointers to code in nilearn & scikit-learn
nilearn.github.io — scikit-learn.org
Use the “API reference” to look up functions
and scroll down for examples of usage

3. 1 Intuitions on machine learning
Adjusting models for prediction
G Varoquaux 2

4. 1 Machine learning in a nutshell: an example
Face recognition
Andrew Bill Charles Dave
G Varoquaux 3

5. 1 Machine learning in a nutshell: an example
Face recognition
Andrew Bill Charles Dave
?G Varoquaux 3

6. 1 Machine learning in a nutshell
A simple method:
1 Store all the known (noisy) images and the names
that go with them.
2 From a new (noisy) images, find the image that is
most similar.
“Nearest neighbor” method
G Varoquaux 4

7. 1 Machine learning in a nutshell
A simple method:
1 Store all the known (noisy) images and the names
that go with them.
2 From a new (noisy) images, find the image that is
most similar.
“Nearest neighbor” method
How many errors on already-known images?
... 0: no errors
Test data = Train data
G Varoquaux 4

8. 1 Machine learning in a nutshell
A simple method:
1 Store all the known (noisy) images and the names
that go with them.
2 From a new (noisy) images, find the image that is
most similar.
“Nearest neighbor” method
How many errors on already-known images?
... 0: no errors
Test data = Train data
G Varoquaux 4

9. 1 Machine learning in a nutshell: intuitions
A single descriptor:
one dimension
x
y
G Varoquaux 5

10. 1 Machine learning in a nutshell: intuitions
A single descriptor:
one dimension
x
y
x
y
Which model to prefer?
G Varoquaux 5

11. 1 Machine learning in a nutshell: intuitions
A single descriptor:
one dimension
x
y
x
y
Problem of “over-fitting”
Minimizing error is not always the best strategy
(learning noise)
Test data = train data
G Varoquaux 5

12. 1 Machine learning in a nutshell: intuitions
A single descriptor:
one dimension
x
y
x
y
Prefer simple models
= concept of “regularization”
Balance the number of parameters to learn
with the amount of data
G Varoquaux 5

13. 1 Machine learning in a nutshell: intuitions
A single descriptor:
one dimension
x
y
Two descriptors:
2 dimensions
X_1
X_2
y
The higher the number of descriptors
the more the trouble
G Varoquaux 5

14. 1 Machine learning in a nutshell: intuitions
A single descriptor:
one dimension
x
y
Two descriptors:
2 dimensions
X_1
X_2
y
The higher the number of descriptors
the more the trouble
The higher the required number of subjects
G Varoquaux 5

15. 1 Testing prediction: generalization and cross-validation
[Varoquaux... 2017]
x
y
x
y
G Varoquaux 6

16. 1 Testing prediction: generalization and cross-validation
[Varoquaux... 2017]
x
y
x
y
⇒ Need test on independent, unseen data
Train set Validation
set
Measures prediction accuracy
sklearn.model_selection.train_test_split
G Varoquaux 6

17. 1 Testing prediction: generalization and cross-validation
[Varoquaux... 2017]
x
y
x
y
⇒ Need test on independent, unseen data
Loop
Test setTrain set
Full data
sklearn.
model_selection.
cross_val_score
G Varoquaux 6

18. 2 Machine learning on rest
fMRI
for population imaging
finding differences between subjects
in functional connectomesG Varoquaux 7

19. From rest-fMRI to biomarkers
No salient features in rest fMRI
G Varoquaux 8

20. From rest-fMRI to biomarkers
Define functional regions
G Varoquaux 8

21. From rest-fMRI to biomarkers
Define functional regions
Learn interactions
G Varoquaux 8

22. From rest-fMRI to biomarkers
Define functional regions
Learn interactions
Find differences
G Varoquaux 8

23. From rest-fMRI to biomarkers
Functional
connectivity
matrix
Time series
extraction
Region
definition
Supervised learning
RS-fMRI
Typical pipeline [Varoquaux and Craddock 2013]
1. Define regions
2. Extract times series
3. Build functional-connectivity matrix
4. Apply supervised machine learning
G Varoquaux 9

24. 2 Defining regions from rest-fMRI
Clustering nilearn.regions.Parcellations
k-means
Fast (in nilearn)
No spatial model
⇒ smooth the data
G Varoquaux 10

25. 2 Defining regions from rest-fMRI
Clustering nilearn.regions.Parcellations
k-means
Fast (in nilearn)
No spatial model
⇒ smooth the data
Ward agglomerative clustering
Recursive merges of clusters
Spatial model constraints merges
⇒ fast
... ... ...
... ...
G Varoquaux 10

26. 2 Defining regions from rest-fMRI
Clustering nilearn.regions.Parcellations
k-means
Fast (in nilearn)
No spatial model
⇒ smooth the data
Ward agglomerative clustering
Recursive merges of clusters
Spatial model constraints merges
⇒ fast
Decomposition models
time
voxels
time
voxels
time
voxels
Y +E · S=
25
N
G Varoquaux 10

27. 2 Defining regions from rest-fMRI
Clustering nilearn.regions.Parcellations
k-means
Fast (in nilearn)
No spatial model
⇒ smooth the data
Ward agglomerative clustering
Recursive merges of clusters
Spatial model constraints merges
⇒ fast
Decomposition models
ICA: nilearn.decomposition.CanICA
seek independence of maps
Sparse dictionary learning:
seek sparse maps
nilearn.decomposition.DictLearning
G Varoquaux 10

28. 2 For connectome prediction [Dadi... 2018]
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
Choice of regions for best prediction?
G Varoquaux 11

29. 2 For connectome prediction [Dadi... 2018]
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
Choice of regions for best prediction?
G Varoquaux 11

30. 2 Region definition: resulting parcellations
Dictionary learning Group ICA
Ward clustering K-Means clustering

31. 2 Region definition: resulting parcellations
Dictionary learning Group ICA
Ward clustering K-Means clustering

32. 2 Region definition: resulting parcellations
Dictionary learning Group ICA
Ward clustering K-Means clustering

33. 2 Time-series extraction
Extract ROI-average signal:
Optional low-pass filter
(≈ .1 Hz – .3 Hz)
Regress out confounds (movement parameters, CSF &
white matter signals, Compcorr, Global mean)
Hard parcellations (eg from clustering)
nilearn.input_data.NiftiLabelsMasker
Soft parcellations (eg from ICA)
nilearn.input_data.NiftiMapsMasker
G Varoquaux 13

34. 2 Connectome: building a connectivity matrix
How to capture and represent interactions?
G Varoquaux 14

35. 2 Connectome: differences across subjects
0 5 10 15 20 25
0
5
10
15
20
25
0 5 10 15 20 25
0
5
10
15
20
25
0 5 10 15 20 25
0
5
10
15
20
25
0 5 10 15 20 25
0
5
10
15
20
25
Correlation matrices
0 5 10 15 20 25
0
5
10
15
20
25
0 5 10 15 20 25
0
5
10
15
20
25
0 5 10 15 20 25
0
5
10
15
20
25
0 5 10 15 20 25
0
5
10
15
20
25
Partial correlation matrices
3 controls, 1 severe stroke patient
Which is which?
G Varoquaux 15

36. 2 Connectome: differences across subjects
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25Large lesion
Correlation matrices
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25Large lesion
Partial correlation matrices
Spread-out variability in correlation matrices
Noise in partial-correlations
Strong dependence between coefficients
[Varoquaux... 2010]
G Varoquaux 15

37. 2 Information geometry: uniform-error parametrization
Subject-specific noise in covariance form manifold
Tangent space removes coupling in coefficients
Controls
Patient
dΣ
M
anifold
Tangent
Tangent embedding[Varoquaux... 2010]
G Varoquaux 16

38. 2 Connectome: which parametrization maps differences?
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25Large lesion
Correlation matrices
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25Large lesion
Partial correlation matrices
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25 Control
0 5 10 15 20 25
0
5
10
15
20
25Large lesion
Tangent-space embedding
[varoquaux 2010]
G Varoquaux 17

39. 2 For connectome prediction [Dadi... 2018]
Time series
2
RS-fMRI
41
Diagnosis
ROIs Functional
connectivity
3
Connectivity matrix
Correlation nilearn.connectome.ConnectivityMeasure
Partial correlations
Tangent space
G Varoquaux 18

40. 2 For connectome prediction [Dadi... 2018]
Time series
2
RS-fMRI
41
Diagnosis
ROIs Functional
connectivity
3
Connectivity matrix
Correlation nilearn.connectome.ConnectivityMeasure
Partial correlations
Tangent space
G Varoquaux 18

41. 2 Supervised learning step [Dadi... 2018]
Functional
connectivity
Time series
3
4
Diagnosis
2
RS-fMRI
1 ROIs
Supervised learning
Stick with Linear models
sklearn.linear_model.LogisticRegression
G Varoquaux 19

42. 2 Supervised learning step [Dadi... 2018]
Functional
connectivity
Time series
3
4
Diagnosis
2
RS-fMRI
1 ROIs
Supervised learning
Stick with Linear models
sklearn.linear_model.LogisticRegression
G Varoquaux 19

43. Predicting from brain activity at rest
RS-fMRI
Functional
connectivity
Time series
2
4
3
1
Diagnosis
ROIs
1. Functional regions (eg clustering, decomposition, or BASC atlas)
2. Filtering and or confound removal
3. Tangent-space parametrization
4. Supervised linear models (eg SVMs)
G Varoquaux 20

44. 3 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.
K. Dadi, M. Rahim, A. Abraham, D. Chyzhyk, M. Milham,
B. Thirion, and G. Varoquaux. Benchmarking functional
connectome-based predictive models for resting-state fmri. 2018.
G. Varoquaux and R. C. Craddock. Learning and comparing
functional connectomes across subjects. NeuroImage, 80:405,
2013.
G. Varoquaux and B. Thirion. How machine learning is shaping
cognitive neuroimaging. GigaScience, 3:28, 2014.
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. 2010.G Varoquaux 21

45. 3 References II
G. Varoquaux, P. R. Raamana, D. A. Engemann, A. Hoyos-Idrobo,
Y. Schwartz, and B. Thirion. Assessing and tuning brain
decoders: cross-validation, caveats, and guidelines. NeuroImage,
145:166–179, 2017.
G Varoquaux 22