Superpixel Segmentation for CC Parcellation in MRI
117
dimension. The CC subdivision is done into an anterior third, the middle of the
anterior and posterior midbody, a posterior third and the posterior one-fifth. The
rostrum, genu, and rostral body presenting the regions of the anterior third illustrate the prefrontal, premotor, and supplementary motor cortical areas. However,
the posterior midbody is crossed by the somaesthesic and posterior parietal fiber
bundles. The sub-regions of the posterior third, containing the isthmus and splenium, are allocated to temporal, parietal, and occipital cortical regions. Thus,
this parcellation method, and as any geometric methods, neither reflects the real
texture nor the internal organization of the CC. In addition, the CC parcellation is strongly dependent on the brain conservation process, since it is based on
post-mortem data. Differently, Hofer proposed the only work based on tractography of DTI (Diffusion Tensor Imaging) by subdividing the CC into five regions
from an average behavior observed via tractography in a specific population of
8 subjects [1]. As already proposed by Witelson, the geometric baseline in the
midsagittal section of the CC is defining the anterior and posterior points of the
structure. The first region, which represents the first sixth, contains fibers projected in the prefrontal region. The remainder of the anterior half CC illustrates
the second region containing the fibers that form the motor and motor areas
of the cerebral cortex. In fact, these fibers form together the largest CC region
and are placed in the back section of the structure. The third region presents
the posterior half minus the posterior third. It contains fibers responsible for
the primary motor cortex. However, this part of the parcellation scheme is in
conflict with Witelson’s method. The fourth region forms third minus the posterior quarter, presenting the primary sensory fibers. The last and the fifth region
represents the CC posterior quarter crossed by the parietal, temporal and visual
fibers. Figure 2 shows a comparison between the geometric schemes proposed by
Witelson and Hofer. We notice that geometric methods allow only to divide the
CC into the same regions among all subjects without considering the human and
individual brain features between different subjects. On the second hand, differently to geometric parcellation methods, Rittner proposed a data-driven method
based on the Watershed technique [15]. This method is composed of four steps.
The first step consists in the weighting of the fractional anisotropy. The second
step performs the selection of the brain midsagittal plane, followed by the third
and the last step which are the CC segmentation using the Watershed technique,
and its parcellation with fixed markers. Nevertheless, this method suffers from
sensitivity to parameters selection. In order to overcome its limitations, Cover
extended the Rittner method with some important changes [12]. Practically, the
author replaced all steps except the first step in order to lead to a more robust
data-driven method. Indeed, the parcellation is improved by applying the Kmeans algorithm after defining the CC centerline. When comparing this method
to that of Rittner, and although both are based on Watershed, it is confirmed
that this method had a better generalization ability using no fixed markers to
execute the Watershed transform. However, due to the lack of quantitative metrics and reference standards, these methods cannot be correctly validated.
117
dimension. The CC subdivision is done into an anterior third, the middle of the
anterior and posterior midbody, a posterior third and the posterior one-fifth. The
rostrum, genu, and rostral body presenting the regions of the anterior third illustrate the prefrontal, premotor, and supplementary motor cortical areas. However,
the posterior midbody is crossed by the somaesthesic and posterior parietal fiber
bundles. The sub-regions of the posterior third, containing the isthmus and splenium, are allocated to temporal, parietal, and occipital cortical regions. Thus,
this parcellation method, and as any geometric methods, neither reflects the real
texture nor the internal organization of the CC. In addition, the CC parcellation is strongly dependent on the brain conservation process, since it is based on
post-mortem data. Differently, Hofer proposed the only work based on tractography of DTI (Diffusion Tensor Imaging) by subdividing the CC into five regions
from an average behavior observed via tractography in a specific population of
8 subjects [1]. As already proposed by Witelson, the geometric baseline in the
midsagittal section of the CC is defining the anterior and posterior points of the
structure. The first region, which represents the first sixth, contains fibers projected in the prefrontal region. The remainder of the anterior half CC illustrates
the second region containing the fibers that form the motor and motor areas
of the cerebral cortex. In fact, these fibers form together the largest CC region
and are placed in the back section of the structure. The third region presents
the posterior half minus the posterior third. It contains fibers responsible for
the primary motor cortex. However, this part of the parcellation scheme is in
conflict with Witelson’s method. The fourth region forms third minus the posterior quarter, presenting the primary sensory fibers. The last and the fifth region
represents the CC posterior quarter crossed by the parietal, temporal and visual
fibers. Figure 2 shows a comparison between the geometric schemes proposed by
Witelson and Hofer. We notice that geometric methods allow only to divide the
CC into the same regions among all subjects without considering the human and
individual brain features between different subjects. On the second hand, differently to geometric parcellation methods, Rittner proposed a data-driven method
based on the Watershed technique [15]. This method is composed of four steps.
The first step consists in the weighting of the fractional anisotropy. The second
step performs the selection of the brain midsagittal plane, followed by the third
and the last step which are the CC segmentation using the Watershed technique,
and its parcellation with fixed markers. Nevertheless, this method suffers from
sensitivity to parameters selection. In order to overcome its limitations, Cover
extended the Rittner method with some important changes [12]. Practically, the
author replaced all steps except the first step in order to lead to a more robust
data-driven method. Indeed, the parcellation is improved by applying the Kmeans algorithm after defining the CC centerline. When comparing this method
to that of Rittner, and although both are based on Watershed, it is confirmed
that this method had a better generalization ability using no fixed markers to
execute the Watershed transform. However, due to the lack of quantitative metrics and reference standards, these methods cannot be correctly validated.