Image Processing Reference
In-Depth Information
Table 2.
Registration results on fetal brain development dataset. The proposed method
yields lowest SSD and comparably smooth deformation fields.
pair-wise uniform adaptive
Mean Error 1
.
36
×
10
4
1
.
36
×
10
4
1
.
31
×
10
4
Smoothness
6
.
12
×
10
3
6
.
72
×
10
3
6
.
2
×
10
3
Fig. 10.
Detail of (from left to right) target image, no, spatially uniform and spatially
adaptive temporal smoothing. Arrows indicate exemplary regions where oversmoothing
is reduced by spatially adaptive temporal smoothing.
4 Discussion
We have presented a method for temporal smoothing of consecutive image regis-
trations in the framework of diffeomorphisms parametrized by stationary veloc-
ity fields. This has enabled us to define a spatially adaptive temporal smoothing
prior that enables the consideration of image information while at simultaneously
resulting in smoother overall deformations. We have successfully evaluated the
proposed methods on synthetic image sequences of simplified cortical folding.
By building deformation models on subsampled datasets, we were able to show
that the temporal smoothing accurately reflects the underlying deformation. We
could replicate these results on a dataset of fetal brain development, yielding a
decrease in registration error and sharper image detail.
In its simplest formulation presented in this paper, the proposed method
can be employed online with only the additional computational cost of just one
interpolation step compared to sequential pair-wise registrations. In cases where
all imaging time-points are available a-priori, more complex forward-backward
smoothing schemes can be envisioned. However, propagating deformation priors
further than one time-step in either direction require to enforce parallelism in
the construction [
29
] and lead to an increase in complexity and thus susceptibil-
ity to registration error. Depending on the nature of the observed process, the
selection of an initial time-point for the smoothing is likely to not only yield
further increases in modeling accuracy but also deepen the understanding of the
underlying dynamics. Evaluation of these effects, notably in the context of brain
growth, will be the focus of future work.
References
1. Scott, J.A., Habas, P.A., Kim, K., Rajagopalan, V., Hamzelou, K.S., Corbett-Detig,
J.M., Barkovich, A.J., Glenn, O.A., Studholme, C.: International Journal of Devel-
opmental Neuroscience. International Journal of Developmental Neuroscience
29
(5),
529-536 (2011)
Search WWH ::
Custom Search