Construction of a 4D Brain Atlas and Growth Model Using Diffeomorphic Registration - Longitudinal Registration and Shape Modeling - page 24

Construction of a 4D Brain Atlas and Growth
Model Using Diffeomorphic Registration

B

Andreas Schuh 1(

) , Maria Murgasova 2 , Antonios Makropoulos 1 ,
Christian Ledig 1 , Serena J. Counsell 2 ,JoV.Hajnal 2 ,
Paul Aljabar 2 , and Daniel Rueckert 1

1 Imperial College London, London, UK
[e-mail]
2 King's College London, London, UK

Abstract. Atlases of the human brain have numerous applications in
neurological imaging such as the analysis of brain growth. Publicly avail-
able atlases of the developing brain have previously been constructed
using the arithmetic mean of free-form deformations which were obtained
by asymmetric pairwise registration of brain images. Most of these atlases
represent cross-sections of the growth process only. In this work, we
use the Log-Euclidean mean of inverse consistent transformations which
belong to the one-parameter subgroup of diffeomorphisms, as it more nat-
urally represents average morphology. During the registration, similarity
is evaluated symmetrically for the images to be aligned. As both images
are equally affected by the deformation and interpolation, asymmetric
bias is reduced. We further propose to represent longitudinal change
by exploiting the numerous transformations computed during the atlas
construction in order to derive a deformation model of mean growth.
Based on brain images of 118 neonates, we constructed an atlas which
describes the dynamics of early development through mean images at
weekly intervals and a continuous spatio-temporal deformation. The evo-
lution of brain volumes calculated on preterm neonates is in agreement
with recently published findings based on measures of cortical folding of
fetuses at the equivalent age range.

1 Introduction

Brain atlases have numerous applications in neurological image analysis. Brain
templates and tissue probability maps are frequently used for image segmenta-
tion [ 9 , 10 ]. Deformations encoding the average brain growth of a population may
be analyzed to study brain development [ 1 ]. Only recently, spatio-temporal (4D)
atlases of the developing human brain have become available: Habas et al. [ 7 ]
created an atlas from 20 fetal Magnetic Resonance (MR) images from polynomial
fits for parameters which describe global scaling, local deformations, and inten-
sity changes. In contrast, Kuklisova-Murgasova et al. [ 9 ] used a non-parametric
kernel regression of ane transformations to build an atlas of the preterm brain.
This has been shown to improve intensity-driven tissue segmentation [ 9 , 10 ]. For