Biomedical Engineering Reference
In-Depth Information
4.1. White Matter Analysis
There is increasing agreement among structural imaging studies on the abnor-
mal anatomy of the white matter in autistic brains. Asymmetry of the WM was
reported by Herbert et al. [44] in a morphometric analysis of postmortem brain
tissue. In addition, volumetric studies of the WM have reported early overgrowth
of WM in autistic children, followed by a reduction in adulthood [17, 44, 28].
In autism, the WM grows normally during the first nine months, then by 2 years
excessive WM is found in some brain areas, such as the frontal lobes and the
cerebellum [44].
Beyond measurement of the volume of WM, diffusion tensor imaging (DTI)
studies have reported that the fractional anisotropy (FA), which provides an index
of structural WM integrity, is reduced in the WM of autistic children relative to
controls. More recently, another DTI study showed that these reductions of FA in
the WM persist into adulthood in several cortical regions of the autistic brain [76].
This coherent finding related to the anomalies in the WM of autistic brains
relative to healthy ones is one of the main motivations behind our study. We intend
to use the difference in WM anatomy in order to develop a classification approach
that separates the two groups based on MRI analyzes.
First, we believe that the reported WM anomalies explain the difference in
folds and gyrifications of the WM between autistic and typical brains, which we
have observed on the data at hand, as shown in Figure 12. Note that the WM
was segmented using our level set-based segmentation technique as explained in
Section 3.2.
For classification purposes, we use the distance map inside of the segmented
WM as a shape representation of the WM structure. We first computed the dis-
tance maps inside the WM of four autistic brains, and that inside the WM of six
healthy brains from the postmortem data, and we did the same for 4 autistic and
16 healthy brains from the savant data. Second, we compute the corresponding
cumulative distribution functions of these distance maps as shown in Figure 13a
for the postmortem data, and in Figure 13b for the savant data. It is clear from
these figures that the two classes, autistic and normal, are completely separable,
which encourages us to use the CDFs of distance maps as a discrimination measure
between the two classes. Furthermore, in order to remove the volume effect, these
CDFs are normalized between 0 and 1, and then averaged, as shown in Figure 14.
Finally, given the brain MRI scan of a subject to be classified, we compute
and normalize, in similar way, the CDF of the distance map inside of its segmented
white matter and compare this CDF to the two average CDFs (autistic and normal)
using the Levy distance. The smallest Levy distance indicates the class to which
the test subject belongs. Figure 15 shows the classification results of two test
subjects from the savant datasets. Both subjects were classified successfully. This
classification method was tested on the two types of datasets. Table 1 summarizes
the performance of the proposed classification method.
