Image Processing Reference
In-Depth Information
Table 1.
AIC comparisons for separate and joint fits of multivariate synthetic data
Relation between growth parameters of variables AIC(Var.1) + AIC(Var.2) AIC(Var.1 + Var.2)
Strong Correlation
-869.751
-1015.242
No Correlation
-866.006
-864.939
The individual and subject specific trends were estimated using NLME based
mixed effects analysis. A univariate NLME model fit was first done separately
for each variable, followed by a joint modeling for both variables using multi-
variate NLME. We first consider (i), the case where the variable parameters are
strongly correlated. Here, the multivariate NLME fit for all variables resulted in
a significantly lower AIC (Akaike Information Criterion) value compared with
the sum of AIC values of univariate fits for each variable. This indicates that
multivariate NLME provides a better fit for the data in case (i). In case (ii) where
the parameters are uncorrelated, the usage of multivariate NLME had no major
effect on the AIC as seen in Table 1. This synthetic data experiment reinforces
the necessity of the multivariate fit for modeling multimodal data, particularly
when correlation exists between modalities.
3.1
Multimodal Contrast Modeling and Analysis on Infant Clinical
Data
The framework outlined in Section 2 is applied to 22 healthy controls scanned
at approximately 6 months, 12 months, and 24 months of age. Registration
removes all volumetric and morphometric differences and segmentation classifies
each voxel into one of the major tissue classes. Intensity distributions for white
and gray matter tissue classes are computed. Four major cortical regions in left
and right hemispheres (eight brain regions in total) are chosen to explore spa-
tially dependent brain maturation patterns. Contrast in T1W and T2W modal-
ities are jointly modeled for each lobe using multivariate NLME. The contrast
value modeled also has a direction attribute (relative to adult-like image) given
by CONTDIR. Results shown are for left hemispheric cortical lobes, although
similar patterns are replicated in the right hemisphere as well.
As seen in Fig.
6
, we infer that contrast change in T1W scans takes place
more rapidly as compared with T2W scans. From visual analysis of the growth
trajectories it is observed that the white-gray contrast in T1W scans becomes
close to adult-like at around 10 months of age. In comparison, contrast in T2W
scans continues increasing until two years of age. Since myelination is known to
be one of the key processes contributing to contrast in T1W and T2W images,
this pattern is in conformity with the well-established knowledge that in general,
changes associated with myelination are apparent earlier and proceed faster on
T1W images than on T2W images [
1
]. The contrast change trajectories of differ-
ent cortical regions are also known to follow the trend of contrast first appearing
in parietal/occipital lobes, followed by temporal and frontal lobes. Quantitative
results from the applied framework are consistent with qualitative radiological
observations : the contrast value is seen to reach early saturation in occipital and
Search WWH ::
Custom Search