Construction of a 4D Brain Atlas and Growth Model Using Diffeomorphic Registration - Longitudinal Registration and Shape Modeling - page 31

3.2 Neonatal Brain Atlas

From the randomly selected subset of neonatal brain MR images, we constructed
an atlas consisting of a time series of mean T2-w template images and corre-
sponding tissue probability maps for the age range 28-44 weeks GA at regular
time points for each week. Exemplary axial slices of these mean images are shown
in Fig. 3 for qualitative assessment.
A comparison of the T2-w atlas template at age 42 weeks GA to the one
created by Serag et al. [ 18 ], based on the arithmetic mean of FFDs, is given by
Fig. 2 . Our method notably captures the cortical folds of the frontal lobe with
more detail, even though more images contributed to the average. As noted
in [ 18 ], Serag et al. use on average 15-19 images per time point for the atlas
construction, whereas given our dataset and a kernel width ˃ =0 . 5 weeks, the
proposed approach uses 19-28 images per time point, where images with kernel
weight below 1% are not considered for the average. It should be noted that we
used a constant kernel width because of the close to uniform age distribution
of our randomly selected subjects, in which case also the subdivision algorithm
used in [ 18 ] would result in a nearly constant kernel width. An adaptive kernel
width may be used in case of a non-uniform sample distribution. It should also
be noted, that the atlas created by Serag et al. was created from a different
subset of available neonatal brain images than the atlas presented in Fig. 3 .

Fig. 3. T2-w templates, cortical grey matter, and white matter maps of our atlas