Automatic Segmentation of Immature Brain MR Images

1
Automatic Segmentation of Immature Brain MR
Images Maria Murgasova, Daniel Rueckert, Leigh
Dyet, James Boardman, David Edwards, Jo
Hajnal Visual Information Processing Group,
Imperial College, London
Motivation Around 5 of all children are born
prematurely and most of them have excellent
chances for survival. However, many of them will
suffer behavioural or learning difficulties (such
as hyperactivity) later in life. For very
immature pre-term infants, abnormal brain
development outside uterus often results in
neurological and psychiatric problems. Current
research tries to understand links between the
brain shape and volume and changes in
neurological and behavioural development in
prematurely born children. High-resolution
magnetic resonance (MR) imaging enables us to
detect these changes. Changes in brain
development can be quantified by measuring volume
of the whole brain, different tissues (white
matter, cortical grey matter, central grey
matter) or various parts of the brain, e.g. basal
ganglia, thalamus, visual cortex. The volume
measurements of different populations of children
will enable clinicians to answer different
questions, such as Does the abnormal white
matter and basal ganglia in neonates results in
abnormal growth later on? Does the proportion of
cortex to white matter follow the rules it
should? To obtain sensitive and precise
measurements, accurate automatic segmentation
methods are necessary. The most common task is to
segment image into white matter, grey matter,
cerebrospinal fluid and non-brain tissues. There
has been a lot of research in past decade towards
automatic techniques for the segmentation of
normal adult brain. However, neonatal and young
children brain MR images have different
properties and special methods are therefore
needed for successful segmentation of these
images.
Method We segmented 52 T1-weighted MR images of
1 and 2 year old children acquired on 0.5 T
Apollo MRI scanner using a method 1-3 based on
Expectation-Maximisation algorithm framework 4.
Intensity values of 4 brain tissues (white
matter, grey matter, cerebrospinal fluid, other
tissues ) were modelled by Gaussian
distributions. A probabilistic brain atlas of 5-9
year old children 5 was aligned with the images
by affine registration and used as prior
information for the EM algorithm. The images were
brainmasked using the aligned atlas. The
algorithm iteratively calculates Gaussian
distribution parameters and segmentation estimate
using the atlas at each iteration until the log
likelihood converges. Our data did not seem to be
strongly influenced by intensity inhomogeneity so
the bias correction step of the method 1 was
not necessary to implement. We also skipped MRF
step as we think it does not bring significant
improvement of the segmentation.
Expectation-Maximization iteratively calculates
the Gaussians and segmentation
Gaussian Mixture Model tissue intensities
modelled by Gaussians
Prior information probabilistic atlas aligned
with data
Segmentation of 1-year-old brain not satisfactory
in central brain structures
Results We obtained reasonable results in the
cortex area of the brain. However, in the central
structures of the brain the segmentation failed.
Large areas of deep grey matter were classified
as white as in young children the intensity
values for deep grey matter are much higher than
for cortical grey matter and cannot be therefore
separated from white matter values on global
scale (Fig. 3). In addition, brainmasking tended
to remove parts of the cortex as most of the
imaged children were premature and their head
shape varied significantly. Therefore the affine
registration was not flexible enough to find the
correct head shape.
Discussion The EM algorithm 1-3 has been
successfully used for segmentation of adult
brains. However it did not prove to satisfactory
segment immature brains. The main reason for its
failure in central brain structures is much
larger natural intensity variation within the
tissues in young developing brains. Another
reason is strong influence of prior information.
Unfortunately, there are no atlases available for
1-2 year old children and standard adult or 5-9
year old atlas are created from about hundred of
subjects by 9-parameter affine transformation and
are therefore very blurry to describe the spatial
position of tissues well which seems to be very
important for young brains.
White and gray matter are more difficult to
separate on global scale in 1-year-old brain than
in adult brain
Future work We would like to improve the
performance of the EM algorithm on immature
brains by creating a more appropriate atlas and
aligning it with the image by non-rigid
registration which will result in better
separation of the tissues in problematic areas of
the brain. This should also solve the problem of
imprecise brainmasking. We will soon obtain one
manual segmentation of 2-year old brain for this
purpose. We will also try to create a deformable
atlas and transfer the labels from the atlas to
images. We expect this approach to work well in
central brain structures which do not vary very
much between the different subjects and produce
worse results in cortical area. Therefore we
might combine both approaches to produce better
segmentations of 1 and 2 year old brains.
(a) histogram of an adult brain image
(b) histogram of 1-year-old-brain image
References 1 K. Van Leemput, F. Maes, D.
Vandermeulen, P. Suetens. Automated model-based
bias field correction of MR images of the brain.
IEEE Trans. Med .Imag. (Special Issue on
Model-Based Analysis of Medical Images), vol. 18,
pp. 885-896, Oct 1999 2 K. Van Leemput, F.
Maes, D. Vandermeulen, P. Suetens. Automated
model-based tissue classification of MR images of
the brain. IEEE Trans. Med .Imag. (Special Issue
on Model-Based Analysis of Medical Images), vol.
18, pp. 897-908, Oct 1999 3 K. Van Leemput, F.
Maes, D. Vandermeulen, P. Suetens. Automatic
segmentation of brain tissues and MR bias field
correction using a digital brain atlas. In Proc.
Medical Image Computing Computer-Assisted
Intervention MICCAI98 (Lecture Notes in
Computer Science). Berlin, Germany
Springer-Verlag, 1998, vol. 1496 pp.
1222-1229 4 A. P. Dempster, N. M. Laird, D. B.
Rubin. Maximum likelihood from incomplete data
via the EM algorithm. J. R. Stat. Soc., vol. 39,
pp. 1-38, 1977. 5 M. Wilke, V. J. Schmithorst,
S. K. Holland. Normative paediatric brain data
for spatial normalisation and segmentation
differs from standard adult data. Magnetic
Resonance in Medcine 50, pp. 749-757, 2003