Automatic Brain and Tumor Segmentation

1
Presence of pathology
Data 3D MRI (1x1x1.5 and 1x1x3 mm3 voxels)
T1 post T1 pre T2
EM segmentation using atlas of normals
Automatic Brain and Tumor Segmentation Nathan
Moon1, Elizabeth Bullitt2,1, Marcel Prastawa1,
Koen van Leemput4,5, and Guido Gerig1,3 1 Dept.
of Computer Science, 2 Dept. of Surgery, and 3
Dept. of Psychiatry, University of North
Carolina, Chapel Hill, NC 27599, USA 4
Radiology-ESAT/PSI, B-3000 Leuven, Belgium 5
Dept. of Radiology, Helsinki University Central
Hospital, Helsinki, Finland
Classification fails in the presence of pathology
NEW METHOD

• Modify normal atlas with individual subjects
  information about pathology
• Tumor model Region of partial enhancement as
  spatial prior and for estimating tumor pdf, low
  probability throughout brain region
• Edema model prior estimate of pdf and
  restriction to white matter region
• Segmentation based on T1 pre-contrast and T2
  channel only (T1 post only for atlas modif.)

ABSTRACT
Atlas Modification for Tumor
p 1.0
p 0.05
Combining image segmentation based on statistical
classification with a geometric prior has been
shown to significantly increase robustness and
reproducibility. Using a probabilistic geometric
model of sought structures and image registration
serves both initialization of probability density
functions and definition of spatial constraints.
A strong spatial prior, however, prevents
segmentation of structures that are not part of
the model. In pathological cases, we encounter
either the presentation of new objects that
cannot be modeled with a spatial prior or
regional intensity changes of existing structures
not explained by the model. Our driving
application is the segmentation of brain tissue
and tumors from multi-channel three-dimensional
magnetic resonance imaging (MRI). Our goal is a
high-quality segmentation of healthy tissue and a
precise delineation of tumor boundaries. We
present an extension to an existing expectation
maximization (EM) segmentation algorithm that
modifies a probabilistic brain atlas with an
individual subject's information about tumor
location obtained from subtraction of post- and
pre-contrast MRI. The new method handles various
types of pathology, space-occupying mass tumors
and infiltrating changes like edema. Preliminary
results on five cases presenting tumor types with
very different characteristics demonstrate the
potential of the new technique for clinical
routine use for planning and monitoring in
neurosurgery, radiation oncology, and radiology.
-

T1 Difference Image
Remap Posterior
Smoothness Constraints (MCF), base prob.
Tumor prior calculated from the T1 pre- and
post-contrast difference image. Mixture-model
fitting gives a posterior probability of contrast
enhancement. Difference intensities are remapped
to a tumor probability image and smoothed (MCF).
Posterior probab. tumor
Multi-parameter fit
Atlas Modification for Edema
p 0.1wm
Edema prior calculated from white matter atlas
channel, based on the observation that edema is
most apparent in white matter. Cluster
initialized between wm and csf clusters in T1/T2
joint histogram.
White Matter Atlas
Edema Prior
PROBLEM
Modified Atlas
wm gm csf
tumor edema
Final spatial prior adds tumor and edema
channels, with normalization of atlas channels to
sum to 1.
Segmentation of medical images in the presence of
pathology presents a significant problem, in
particular due to high variability in the size,
shape, and appearance of the pathology. Tumor
segmentation is highly relevant for diagnosis and
monitoring, surgery, and for studying
relationships between vascularity and tumors in
surgical planning of embolization and
radiotherapy.
EM Iterations
Progress of tissue density distributions in joint
histogram (top) and of classification (bottom).
it 1 it 2
it 5 it 20
MODELS FOR TUMOR SEGMENTATION

• Radiologist
• Something wrong is there
• Enhancement with contrast agent
• Lump
• Multi-channel intensities
• Deformation of normal anatomy
• Asymmetry

• Models for Image Analysis
• Probabilistic brain atlas, prior for normal
• Use post-pre contrast difference
• Blob Model
• Multivariate statistical analysis
• Atlas Deformation
• Left-Right Deformation

RESULTS
T2
T1 pre
3D view
T1 post

• Method has been applied to 5 cases
• T1pre/T2 channels used for classification
• T1 post (enhanced) helps initializing tumor
  cluster prior
• Partial volume errors due to coarse T2 resol.

classification
T2
T1 pre
classification
3D view
T1 post
EXISTING BRAIN SEGMENTATION
Existing framework for automatic segmentation of
healthy brain tissue from head MRI Koen van
Leemput, et. al., KUL 1
T2
T1 pre
classification
3D view
T1 post
bias estimation

• Expectation Maximization EM algorithm (see also
  Wells et al. 2)
• Single or multiple MRI channels
• Built in bias inhomogeneity correction
• Initialization and classification governed by
  geometric prior (see also Warfield 3, here
  using statistical atlas (ICBM150 4)
• Built in brain stripping

CONCLUSIONS

• Automatic classification of healthy tissue and
  pathology without user interaction
• New approach Modifies normal statistical atlas
  with individual pathology
• Uses spatial priors, intensity model priors, and
  multivariate data
• Future work
• Higher resolution triple channel MRI (3T scanner,
  1x1x1mm3, T1 and T2)
• Integrate high-dimensional warping to include
  atlas deformation and LR asymmetry
• Validation against experts segmentation (Sean Ho
  et al., ICPR 2002 5)

• Expectation step

• Maximization step

• References
• 1 van Leemput, K., Maes, F., Vandermeulen, D.,
  Suetens, P. Automated model-based tissue
  classification of MR images of the brain. IEEE
  TMI 18 (1999) 897-908.
• 2 W. M. Wells, W.E.L. Grimson, R. Kikinis, and
  F.A. Jolesz, Adaptive Segmentation of MRI Data,
  IEEE TMI 15(4)429-443, 1996.
• 3 S. Warfield, M. Kaus, F.A. Jolesz, and R.
  Kikinis, Adaptive template moderated spatially
  varying statistical classication, in MICCAI'98,
  Oct 1998, vol. 1496 LNCS, pp. 431-438, Springer.
• 4 ICBM Atlas, McConnell Brain Imaging Centre,
  Montréal Neurological Institute, McGill
  University, Montréal, Canada.
• 5 Sean Ho, Elizabeth Bullitt, Guido Gerig,
  Level Set Evolution with Region Competition
  Automatic 3-D Segmentation of Brain Tumors, Proc.
  ICPR, IEEE, August 2002

Key idea Probabilistic brain atlas 4 is
registered to subject image to provide spatial
prior. The atlas initializes the distributions
and acts as a prior during classification
ICBM150 MRI and tissue atlas
September 2002