Transcript Fluid Deformation (FD)

Volumetric vs. Surface-based Intersubject Alignment for Localization of Auditory Cortex Activation
Rutvik Desai, Einat Liebenthal, Edward T. Possing, and Jeffrey R. Binder
Department of Neurology, Medical College of Wisconsin, Milwaukee, WI
Introduction
Intersubject Alignment
Results
Discussion
•Brain structures vary in their size, shape, position, and relative
orientation. This renders pooling data across different anatomies, a
common task for group analysis of neuroimaging data, non-trivial.
•In order to compare and overlay anatomies and functional
activations, a one-to-one mapping needs to be specified so that each
location in one brain corresponds to a unique location in another
brain. A variety of alignment techniques address this problem.
Talairach: Individual subject anatomies and functional maps were
projected into the standard stereotaxic space defined by Talairach and
Tournoux (1) using AFNI. Since smoothing of functional data to
account for anatomical variation is a common practice when using
Talairach transformations, a smoothed version (TLB) of the
functional data was also created by applying a Gaussian filter with
FWHM of 4mm. No intensity or cluster thresholds were applied. The
functional volumes were then projected onto the surface model of a
canonical (N27) brain, using the AFNI program 3dVol2Surf,
which takes into account all the voxels along the segments
connecting the white matter and the pial surfaces.
Automated Spherical Warping (ASW): Anatomical surface models
were created for each subject and the N27 brain using FreeSurfer.
The inflated white matter surface was transformed into a sphere, and
warped to a spherical representation of an averaged brain (2). The
functional data was projected onto the N27 brain using the same
transformations..
Fluid Deformation (FD): A patch of left perisylvian regions,
including Heschl’s gyrus (HG), the superior temporal gyrus (STG)
and sulcus (STS), the planum temporale (PT), the middle temporal
gyrus (MTG), the Sylvian fissure (SF), and the inferior frontal gyrus
(IFG) was cut from the inflated surface and flattened. Four landmarks
(outlining HG, STG, MTG, and IFG) were selected as anatomical
markers on each patch, as shown below. These patches were
deformed using fluid deformation (3) and aligned with the N27 brain
using Caret (4).
Phonemes
• The accuracy of the group maps can be evaluated by relating them
Alignment Methods
• Talairach normalization (TL): Volumetric method based on linear
transformations, common, and easy to use (1). Relatively poor
accuracy (~10mm to 15mm).
• Nonlinear volumetric warping: Matches voxel intensity in volumes.
Not based on explicit sulcal and gyral landmarks.
•Automated Spherical Warping (ASW): Creates explicit
representation of cortical surface and transforms it to a spherical
form. The individual sphere is warped match a sphere representing an
average folding pattern (2).
• Fluid Deformation (FD): The cortical surface is flattened and
modeled as a viscoelastic fluid sheet. The sheet is deformed so as to
align the manually selected landmarks to the corresponding
landmarks on a canonical or atlas brain.
Aims
•Empirically evaluate the TL, ASW, and FD alignment techniques for
activation in auditory cortex.
•Obtain high-resolution maps of the areas activated by categorical
perception of speech sounds.
TL
TLB
FD
ASW
Tones
TL
TLB
Conclusions
Methods
The cut and flattened patch
Subjects, stimuli, and Task
Data is reported for 18 right-handed subjects. There were four task conditions as
shown below. The ABX task, consisting of 2-alternative forced-choice
discrimination (is X identical to the first or second token in the preceding AB
pair?), was performed with Phonetic (P) and Tone (T) stimuli. The phonetic test
items consisted of tokens chosen from an 8-token continuum from /ba/ to /da/
(tokens ba1-ba8). The continuum was created by varying the second formant
(F2) continuously between the anchor points /ba/ and /da/. The ba2-ba4 pair
represents a within-category discrimination, while the ba4-ba6 pair involves an
across-category discrimination.
Phonemes
• Surface-based alignment methods provide a superior alternative to
the volume-based Talairach alignment, at least with respect to the
activation in auditory cortex.
• Both ASW and FD can be used to align individual maps at a high
resolution, though FD may be somewhat superior for aligning
relatively minor anatomical features.
• Both ASW and FD are significantly more time- and computationintensive compared to TL.
• ASW requires no user-interaction, while landmarks need to be
selected manually for FD. However, FD may prove to be less time
consuming if one is able to identify a ROI a priori, because FD does
not require spherical transformation and deformation of the entire
hemisphere.
• Human cortical areas engaged in the analysis of speech phonemes
included the STG, the upper bank of the STS, and the lateral aspect
of HG in the left temporal lobe. The activation of the upper bank of
the STS is in accordance with physiological studies in non-human
primates suggesting multimodal functions for this region (5).
ASW
FD
Condition
Baseline
Noise
Pure Tones
to the individual subject maps, where the activation is mapped to the
subject’s own anatomy.
• In the P > T contrast, the TL and TLB methods show a large focus
on the central MTG, while the surface-based methods do not. Visual
inspection of the individual subjects’ data reveals that most subjects
do not show significant foci on the MTG. In TL and TLB, the MTG
focus results from mis-registration due to differences in the precise
location of STG and MTG in various anatomies.
• In the P and T conditions, FD and ASW produce similar maps.
However, FD results in a higher activation of the HG than ASW. The
stronger activation of HG is consistent with the individual subject
maps, and is presumably due to superior alignment when landmarks
are selected manually. Automated methods such as ASW rely on
averaged brains which preserve large-scale anatomical features, but
may not be able to precisely align relatively smaller landmarks such
as HG.
• In P > T, both ASW and FD result in the activation of anterior and
middle STG, the upper bank of the STS, and the lateral aspect of HG.
There is also a small focus on the lower bank of the anterior STS
extending into the MTG. This may potentially be due to “leaking” of
the STG activation across the banks of STS in some subjects, due to
the 4mm slice thickness, and needs further examination. The focus at
the junction of STG and HG is somewhat stronger in FD, while the
focus in the posterior portion of the middle STS is somewhat stronger
in ASW.
Stimuli
Silence
White noise
Low:1000-1008 Hz
High:1008-1016Hz
Within:ba2-ba4
Across:ba4-ba6
Task
Passive listening
Passive listening
ABX
ABX
Imaging Parameters
•Image acquisition: 1.5T GE Signa scanner.
•Functional images: gradient-echo, echo-planar images, clustered acquisition at
TR = 8 sec, TE = 40 ms, Flip angle = 90°, acquisition time = 2200 ms. 22 axial
slices, 3.75 x 3.75 x 4 mm3
• Structural Images: 3-D spoiled gradient-echo sequence. Whole brain sagittal
slices, 0.9 x 0.9 x 1.2 mm3 . Three volumes were co-registered and averaged to
improve image resolution.
• Image analysis: Spatial co-registration, multiple linear regression (AFNI
3dDeconvolve) with reference functions representing the conditions.
Atlas patch with
landmarks
Phonemes > Tones
TL
For all methods with the exception of TLB (which was smoothed
separately), a small amount of smoothing was applied on the surface
functional maps by averaging the activation of each node and its
neighbors. A random-effects analysis was then carried out on the
surface by comparing individual activation maps with a constant
value of 0 on a node-wise basis. The resulting t-maps were
thresholded at node-wise P < 0.001, and clusters smaller than 75mm2
were removed.
An individual patch
TLB
Acknowledgements
We thank David van Essen, John Harwell, Donna Hanlon, Ziad Saad, Rick Reynolds, Brenna
Argall, and Doug Greve for assistance with many technical issues.
ASW
FD
deformed to match the atlas
10-5
10-4
10-3
References
(1) Talairach, J., and Tournoux, P. (1988). Coplanar stereotaxic atlas of the human brain. Thieme
Medical, New York.
(2) Fischl, B. et al. (1999). High-resolution intersubject averaging and a coordinate system for the
cortical surface. Human Brain Mapping, 8:272-284.
(3) Van Essen, D.C. et al. (1998). Structural and functional mapping of human cerebral cortex:
Solutions are in the surfaces. Proc. Natl. Acad. Sci., 95:788-795.
(4) Van Essen, D.C. et al. (2001). An integrated software system for surface-based analyses of
cerebral cortex. Journal of American Medical Informatics Association, 8:443-459.
(5) Poremba, A. et al. (2003). Functional mapping of the primate auditory system. Science,
299(5606):568-572.