Comparison of MM-LDDMM, CARET, FreeSurfer
Cortical Surface Mappings
Jidan Zhong1, Anqi Qiu2
Introduction
Cortical surface-based mapping has been widely used to compensate for the
individual variability of cortical shape and topology in anatomical and
functional studies. While several surface mapping methods were proposed
using different features or similarity metrics, a few studies have extensively
and quantitatively evaluated different surface mapping methods. In this study
we compared three algorithms for mapping surfaces, including multimanifold large deformation diffeomorphic metric mapping (MMLDDMM)[1], FreeSurfer (v4.0.2)[2] , and CARET (v5.61)[3].
Methods
The cortical surfaces of 40 subjects (20F, mean age: 54y) and one template
(F, age: 54y) were generated by FreeSurfer. We semi-automatically
delineated 14 sulcal and 12 gyral curves (see the above figure) on each
surface and applied MM-LDDMM to simultaneously map the curves and the
cortical surface to those of the template. After the spherical parameterization
of the cortical surface, each subject sphere was mapped to the template
sphere by both FreeSurfer and CARET methods, where FreeSurfer aligned
cortical surfaces by maximizing the similarity of their cortical convexity and
CARET registered cortical surface using six landmark curves.
Results
6.36 (±3.31) mm2 for MM-LDDMM, 14.38 (±6.77) mm2 for FreeSurfer, and
14.88 (±9.19) mm2 for CARET.
Curvature correlation : The curvature correlation between the subject’s
deformed cortical surface and the template surface was calculated to investigate
whether local and global shape characteristics of the two surfaces were similar
after mapping. The Student t-tests on the correlation scores revealed the accuracy
of the curvature pattern alignment decreasing in order of using the MM-LDDMM
(0.50±0.03), FreeSurfer (0.34±0.04), and CARET (0.30±0.03) algorithms.
Local deformation error: To directly quantify
deformation that each mapping algorithm
characterizes, we generated simulated surfaces
from a template with known deformation as
ground truth. We modeled the random
deformation field as a linear combination of the
Laplace-Beltrami (LB) bases ψ(x)[5] (see the left
figure) on the template surface in the form of
r
n
r
r
U ( x ) = U 0 + ∑ U iψ i ( x ) N ( x )
i =1
r
where N (x) was the normal vector at location x on the template surface, U 0 and {Ui }in=1
were Gaussian distributed random variables. Ten surfaces were generated based
on the first ten LB bases and mapped to the template with the three methods. The
local deformation error maps averaged over the ten simulated datasets and their
distribution for each of the mapping algorithms were shown in the figure below.
Surface alignment consistency (SAC): The SAC[3] was computed for 17
sulcal regions (see the above figure) distributed broadly over the cortical
surface. It was introduced for quantifying the anatomical variability of a
sulcal region among a group of subjects that can be characterized by the
cortical mapping algorithms. The figure below shows SAC measures of each
sulcal region for the three methods. Overall, the SAC values averaged across
all regions were significantly higher for MM-LDDMM than FreeSurfer and
CARET (p<0.05). The FreeSurfer and CARET mappings performed
equivalently in terms of the SAC value.
Curve variation error : The Hausdorff distances between curves
transformed to the template by the three methods were computed for each
sulcal/gyral curve. Then the curve variation error[4] (see the top right figure)
was calculated to evaluate the anatomical variation of a specific curve among
subjects. The curve variation errors averaged over all twenty six curves were
1 NUS Graduate School for Integrative Sciences and Engineering , National University of
Singapore, Singapore
2 Division of Bioengineering, National University of Singapore, [email protected]
Computational Functional Anatomy Laboratory
http://www.bioeng.nus.edu.sg/cfa/index.html
Conclusion
We quantitatively evaluated the accuracy of three surface mapping algorithms
(MM-LDDMM, FreeSurfer, and CARET) using 40 MRI scans and 10 simulated
datasets. We demonstrated that each method performed well in its own similarity
metric. Overall, the MM-LDDMM method provides better mapping accuracy
because it aligns local curves and the global surface simultaneously. CARET
shows better local accuracy than FreeSurfer around the regions with landmarks
which, however, can not control the global shape well, while FreeSurfer shows
better global accuracy but lacks local features to constrain the local shape.
References
[1] Zhong, J., Qiu, A., NeuroImage, 49: 355-365, 2010.
[2] Fischl et al, Hum Brain Mapp, 8:272-284, 1999.
[3] Van Essen, NeuroImage, 28:635-662, 2005.
[4] Pantazis et al., NeuroImage, 49:2479-2493, 2010.
[5] Qiu A, et al., IEEE Trans Med Imaging, 25:1296-1306, 2006.