clustering cortical searchlights based on shared...
TRANSCRIPT
http://www.pymvpa.org
Clustering cortical searchlights based on shared representational geometry
Matteo Visconti di Oleggio Castello1,*, Samuel A. Nastase1,*,James V. Haxby1,2, M. Ida Gobbini1,3, Yaroslav O. Halchenko1
1Dept. of Psychological and Brain Sciences, Dartmouth College, NH, USA 2Ctr. for Mind/Brain Sci. (CIMeC), Univ. degli Studi di Trento, Rovereto, Italy 3Dipartimento di Medicina Diagnostica e Sperimentale, Medical School, Univ. di Bologna, Italy *Equal contribution
Simulation resultsAgainst Ground TruthAgainst PredictionTrue k
ARI AMI Instability Correlation
�
�� ��
�
���
� �
�
�
�
�
�
�
0.00
0.25
0.50
0.75
1.00
com
plete
�
�
��
�����
�
��
�
�
� �
����
����� ��
��
��
�
����
����� ��
��
�
���
� ��
0.00
0.25
0.50
0.75
1.00
km
eans
��
�
�
�
�
���
��
���
����
�
����
���
��
��
����
�
����
��
��
� � ���
�
��
0.00
0.25
0.50
0.75
1.00
gm
m�
tie
d
�
�
�
�
��
�����
�
�
���
�
��
�����
�
��
���
�
��
�����
�
�
�
��
� ����� ��
�
�
0.00
0.25
0.50
0.75
1.00
wa
rd�
unstr
�� �
�
�� �
�
� � �
�
��� ��� ���� ��
0.00
0.25
0.50
0.75
1.00
wa
rd�
str
2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10
k
�
�
�
����
���
�
�
�
�
��
�
�
�
����
��� ���
�
��
�
�
�
����
���
�
��
���
�� ����
�
0.00
0.25
0.50
0.75
1.00
gm
m�
diag
Va
lu
e
AimsApply clustering to searchlight RSA in order to functionally parcellate the cerebral cortex
Investigate reproducibility for different clustering algorithms [1]
Investigate effects of hyperalignment [2]
Compare experiment-specific parcellations to publicly available functional and anatomical parcellations
MethodsClustering algorithms - k-means- Ward (with and without structural constraints)- Gaussian mixture models (with different covariance structures)- Complete linkage with correlation distance
Metrics- Reproducibility procedure (inspired by [3] and [4], see below) using adjusted Rand index (ARI), adjusted mutual information (AMI), instability [3], and correlation of average RDMs between corresponding clusters
- Consistency of representational geometry quantified using a measure of homogeneity [5] based on pairwise correlation distance between all searchlight RDMs within a parcel
S1 S2 S3 S4 S5 S6
Train Test
Clust Train Clust Test
Predict Clust Test
kNN or Prediction
ARIAMI
InstabilityCorrelation
Cro
ss-Valid
atio
n
Repeat fo
r k=2
...K
Repeat fo
r N fo
lds
S1 S2 S3 S4 S5 S6
Cross-validation scheme for parcellation reproducibility estimate
Cluster Ground Truth
- 10 simulated subjects- 6 contiguous clusters with pattern information- Different RDM in each cluster- Random subject-specific and common noise
Simulated data Real fMRI data
- 12 participants watching 2 s naturalistic video clips of behaving animals [6]- 5 animal taxa, 4 actions (20 total conditions)- 1-back task requiring attention to action categories- 20 x 20 RDM computed with surface-based searchlights (100 voxels) using correlation distance- Anatomical alignment and whole-brain hyperalignment [2]
ConclusionsMeaningful parcellations can be obtained by clustering shared representational geometries
None of the methods tested provided "the ultimate solution" for whole brain parcellation with respect to reproducibility
Hyperalignment had drastic effects on reproducibility
Experiment-specific parcellations exhibit higher homogeneity compared to resting-state and anatomical parcellations
Future directions:
Investigate full range of possible values of k
Introduce parcel pruning to remove non-informative parcels
www.github.com/mvdoc/reprclustFork it and try your favorite methods!
[WIP] Code
References[1] Thirion, B., Varoquaux, G., Dohmatob, E., & Poline, J. B. (2014). Which fMRI clustering gives good brain parcellations? Frontiers in Neuroscience.[2] Haxby, J. V., Guntupalli, J. S., Connolly, A. C., Halchenko, Y. O., Conroy, B. R., Gobbini, M. I., Hanke, M., & Ramadge, P. J. (2011). A common, high-dimensional model of the representational space in human ventral temporal cortex. Neuron.[3] Lange, T., Roth, V., Braun, M. L., & Buhmann, J. M. (2004). Stability-based validation of clustering solutions. Neural Computation.[4] Yeo, B. T., Krienen, F. M., Sepulcre, J., Sabuncu, M. R., Lashkari, D., Hollinshead, M., Roffman, J. L., Smoller, J. W., Zöllei, L., Polimeni, J. R., Fischl, B., Liu, H., & Buckner, R. L. (2011). The organization of the human cerebral cortex estimated by intrinsic functional connectivity. Journal of Neurophysiology.[5] Gordon, E. M., Laumann, T. O., Adeyemo, B., Huckins, J. F., Kelley, W. M., & Petersen, S. E. (2014). Generation and evaluation of a cortical area parcellation from resting-state correlations. Cerebral Cortex.[6] Nastase, S. A., Connolly, A. C., Oosterhof, N. N., Halchenko, Y. O., Guntupalli, J. S., Visconti di Oleggio Castello, M., Gors, J., Gobbini, M. I., & Haxby, J. V. (2015). Attention reshapes distributed population codes for semantic representation. Manuscript in preparation.
Homogeneity analysis
Within each parcel, average pairwise correlation distance between all searchlight RDMs
Null distribution of homogeneities estimated by randomly rotating the spherical projection of the cortical surface
Comparison with parcellations derived from anatomy (FreeSurfer) and resting-state functional connectivity [4]
Effect of hyperalignment
ARI AMI Instability
0.0
0.2
0.4
0.6
0.8
wa
rd�
str
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
k
Va
lu
e
Alignmentanat
hyper
Posterior left hemisphere
Whole brain
fMRI resultsARI AMI Instability
�
�
�
�
��
�
��
�
��� �����
�
�
� �
�
��
�
���
�
��
�
�
�
�
�
��
��
��
�
�
�
�� ���
�� �
�� �������
����
��
�����
���
���
� � �
��
��
�
��� �
�
� �������
� �� �
��
��
��
�
�
�
�
�� �
���
�
��
��
�
��
� � � � � � � ��
��
�
�
�
�
�
�
��
�
��
�
�
�
�
�
�
���
��
�
�
�
�
�
���
�
��� ��
�
�
��� ��
�
��
�
�
�
��
�
�
��
��
�
������
�� �
� ��
�
���
�
��
0.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
0.00
0.25
0.50
0.75
gm
m�
sp
hkm
ea
ns
wa
rd�
str
2 4 6 8 1012141618202224262830 2 4 6 8 1012141618202224262830 2 4 6 8 1012141618202224262830
k
Va
lu
e
k-means
k = 2
k = 3
k = 4
Ward (structural constraints)
k = 2
k = 3
.http:// .net
accelerated by
Visit Exhibit Table