[SfN] Agreement between the white matter connectivity via tensor-based morphometry and the volumetric white matter parcellations
1. Nanosymposium: Data Analysis and Statistics
Agreement between the white matter
connectivity based tensor-based
morphometry and the volumetric white
matter parcellations
16 Nov 2011 Wed, 08:00~08:15
Seung-Goo KIM
Department of Brain and Cognitive Sciences,
Seoul National University, Korea (ROK)
2. ACKNOWLEDGEMENT
• Hyekyoung Lee @ SNU
• Moo K. Chung, Jamie L. Hanson,
Richard J. Davidson, Seth D. Pollak
@ U Wisconsin-Madison
• Brain B. Avants, James C. Gee @ U Penn
5. Connectivity based on correlation
• Correlation of functional measures
i j i j i j
scan 1 scan 2 scan 3
6. Connectivity based on correlation
• Correlation of functional measures
i j i j i j
scan 1 scan 2 scan 3
7. Connectivity based on correlation
• Correlation of functional measures
i j i j i j
scan 1 scan 2 scan 3
• Correlation of anatomical measures
i j i j i j
subject 1 subject 2 subject 3
8. Brain network based on
cortical thickness
Worsley et al., 2005, Phil.Trans. R. Soc. B
10. White matter connectivity
• Cortical thickness is only defined along the
cortical surface, thus the connectivity
within the white matter cannot be known.
11. White matter connectivity
• Cortical thickness is only defined along the
cortical surface, thus the connectivity
within the white matter cannot be known.
• We can build the white matter connectivity
using tensor-based morphometry (TBM)
that quantifies local volume at all voxels.
12. TBM-based networks
Kim et al., 2011, IEEE International Symposium on Biomedical Imaging (ISBI), pp. 808-811.
15. OBJECTIVES
• Agreements between TBM-based networks
and DTI-based connectivity
• Differences between a clinical group and a
normal control group using TBM-based
networks
17. Subjects & images
• PI (Post-Institutionalized) subjects
• 32 children who experienced maltreatment
in the early stages of life (<2 yr-old) in
orphanages and were adopted later
• NC (Normal Control) subjects
• 33 age & gender matched children
• T1-weighted MRIs (3 Tesla; 1mm3 voxel)
• Non-linear normalization by ANTS (U Penn)
23. Partial correlation
• To factor out age and gender, first fit a GLM
JD = β0 + β1 · age + β2 · gender + noise
24. Partial correlation
• To factor out age and gender, first fit a GLM
JD = β0 + β1 · age + β2 · gender + noise
• Take residuals of the fit from the observation
r = JD − (β0 + β1 · age + β2 · gender)
25. Partial correlation
• To factor out age and gender, first fit a GLM
JD = β0 + β1 · age + β2 · gender + noise
• Take residuals of the fit from the observation
r = JD − (β0 + β1 · age + β2 · gender)
• Pearson’s correlation between the residuals is
the partial correlation
28. DTI-based white matter atlas
(ICBM-DTI-81)
S. Mori et al., 2008, NI. Ck (k = 1, · · · , 48)
29. Connectivity Matrix
[Xmn ] ∈ R
48×48
i∈Cm ,j∈Cn corr(i, j)
Xmn =
Number of pairs
C4
C3
30. Connectivity Matrix
[Xmn ] ∈ R
48×48
i∈Cm ,j∈Cn corr(i, j)
Xmn =
Number of pairs
C4
C3
31. Connectivity Matrix
[Xmn ] ∈ R
48×48
i∈Cm ,j∈Cn corr(i, j)
Xmn =
Number of pairs
C4
C3
32. WITHIN- vs. BETWEEN-
[Xmn ] ∈ R 48×48
i∈Cm ,j∈Cn corr(i, j)
Xmn =
Number of pairs
m=n m = n
C4
C3 C3
Diagonal element Off-diagonal element
in [Xmn ] in [X
mn ]
33. WITHIN- vs. BETWEEN-
[Xmn ] ∈ R 48×48
WITHIN-
corr(i, j)
i∈Cm ,j∈Cn
connectivity Xmn =
Number of pairs
m=n m = n
C4
C3 C3
Diagonal element Off-diagonal element
in [Xmn ] in [X
mn ]
34. WITHIN- vs. BETWEEN-
[Xmn ] ∈ R 48×48
WITHIN- BETWEEN-
i∈Cm ,j∈Cn corr(i, j)
connectivity Xmn =
Number of pairs connectivity
m=n m = n
C4
C3 C3
Diagonal element Off-diagonal element
in [Xmn ] in [X
mn ]
35. WITHIN- vs. BETWEEN-
[Xmn ] ∈ R 48×48
WITHIN- BETWEEN-
i∈Cm ,j∈Cn corr(i, j)
connectivity Xmn =
Number of pairs connectivity
m=n m = n
C4
C3 C3
Diagonal element Off-diagonal element
in [Xmn ] in [X
mn ]
36. WITHIN- vs. BETWEEN-
[Xmn ] ∈ R 48×48
WITHIN- BETWEEN-
i∈Cm ,j∈Cn corr(i, j)
connectivity Xmn =
Number of pairs connectivity
m=n m = n
C3
Diagonal element
C3
C4
Off-diagonal element
in [Xmn ] in [X
mn ]
37. Statistical inferences
• Jackknifing on NC and PI
• 500 random networks as null models
against brain networks
• Willcoxon rank sum test is used
44. NC vs. PI: Local inference
p0.01, Bonferroni corrected
NC PI
PINC, PINC
45. NC vs. PI: Local inference
p0.01, Bonferroni corrected
NC PI
PINC, PINC
46. NC vs. PI: Local inference
p0.01, Bonferroni corrected
NC PI
PINC, PINC
47. NC vs. PI: local inference
PINC PINC
GCC: Genu of corpus callosum EC-R: External capsule, right
SCR-L: Superior corona radiata, left FX/ST-R: Fornix / Stria terminalis, right
SCP-L: Superior cerebellar peduncle, left
49. Conclusions
• The greater within-connectivity than
between-connectivity in brain networks
shows agreement between TBM-based
network and DTI-based atlas.
50. Conclusions
• The greater within-connectivity than
between-connectivity in brain networks
shows agreement between TBM-based
network and DTI-based atlas.
• Locally differences in within-connectivity
may imply altered white matter integrity
due to early maltreatment