Automatic Segmentation of Neonatal Brain MRI

1
Automatic Segmentation of Neonatal Brain MRI

• Marcel Prastawa1, John Gilmore2, Weili Lin3,
  Guido Gerig1,2
• University of North Carolina at Chapel Hill
• 1Department of Computer Science
• 2Department of Psychiatry
• 3Department of Radiology
• Partially supported by
• NIH Conte Center MH064065 and NIH-NIBIB R01
  EB000219

2
Goal

• Segmentation of brain tissues of newborn infants
  from multimodal MRI
• Particular interest in the developing white
  matter structure
• Motivation
• Analysis of growth patterns, study of
  neuro-developmental disorders starting at a very
  early age

csf
nWM
gm
mWM
3
Imaging the Developing Brain
age
35 weeks
44 weeks
15 months
2 years
adult
Rhesus equiv. 6yrs
4
Challenges

• Smaller head size
• Low contrast-to-noise ratio
• Intensity inhomogeneity
• Motion artifacts
• Division of white matter into
• myelinated and non-myelinated
• regions
• Previous work
• Warfield et al 1998 (methodology)
• Hüppi et al 1998 (clinical study)

T1
T2
Labels
csf
nonm. WM
gm
myel. WM
5
Challenges
Neonate 2 weeks
Adult
CNR 2.9
CNR 6.9
6
Approach

• Non-optimal input data, rely on high level prior
  knowledge
• Intensity ordering (e.g. in T2W)
• wm-myelinated lt gm lt wm-non-myelinated lt csf
• Aligned spatial priors (brain atlas)
• White matter is considered as one entity

7
Method Overview
Segmentation - Bias correction
Initialization
Refinement
8
Intensity Clustering

• Samples obtained by thresholding atlas priors
• T1 T2 Pr(wm, x)
  Overlay
• Noisy data, low contrast ? robust techniques
• Two robust estimation techniques
• Minimum Spanning Tree (MST) clustering
• Minimum Covariance Determinant (MCD) estimator
• Obtain initial estimates of intensity
  distributions

9
Minimum Spanning Tree Clustering

• Cocosco et al 2003
• Break long edges in MST, example
• Detect multiple clusters while pruning outliers
• Iterative process, stops when cluster feature
  locations are in the desired order
• Feature location summary value of cluster
  intensities

10
Determining Feature Locations

• Need reliable location estimate to find good
  clusters
• Standard estimates (e.g., mean, median) not
  always optimal
• Use robust estimator to determine location of a
  compact point set in a cluster

Median
Mean
11
Minimum Covariance Determinant

• Rousseeuw et al 1999
• Feature location of MST clusters to determine
  ordering?
• Smallest ellipsoid that covers at least half the
  data
• MCD gives robust location estimate
• Example

12
Intensity Clustering Algorithm

• Apply MCD to GM and CSF samples obtain T2
  locations
• Construct MST from WM samples
• T ? 2
• Repeat until T 1
• Break edges longer than T x (local average
  length)
• Find largest myelinated WM cluster, where
• T2myel lt T2GM
• Find largest non-myelin. WM cluster, where
• T2GM lt T2non-myel lt T2CSF
• Stop if WM clusters found
• Otherwise, T ? T 0.01

13
Method Overview
Segmentation - Bias correction
Initialization
Refinement
Initial intensity Gaussian PDFs
14
Bias Correction

• Wells et al 1996, van Leemput et al 1999
• Bias RF inhomogeneity and biology
• Images low contrast, histogram is smooth
• Use spatial context, bias is log-difference of
  input intensities and reconstructed flat image
• Fit polynomial to the bias field (weighted least
  squares)
• Interleaves segmentation and bias correction

Gaussian intensity PDFs
15
Method Overview
Segmentation - Bias correction
Initialization
Refinement
Bias corrected images Segmentations
16
Refinement

• Previous stage assumes Gaussian intensity
  distributions
• May have non-optimal decision boundaries due to
  overlap
• Re-estimate intensity parameters from
  bias-corrected images
• MST clustering to obtain training data
• Parzen windowing to estimate density

Parzen kernel density estimate
Atlas prior
17
Results 1/2

• UNC Radiology Weili Lin (Siemens 3T head-only)
• UNC-0094
• UNC-0096

Classification
T1
T2
3D
Classification
T1
T2
3D
18
Results 2/2

• Provided by Petra Hüppi (Geneva, Philips 1.5T)
• Geneva-001
• Geneva-002

T2
Classification
T1
3D
T2
Classification
T1
3D
19
Results UNC 0096
Upper row T1, T2w, Tissue labels, registered
atlas Lower row Probabilities for wm-myel, wm,
gm, csf
20
Results UNC 0096
21
Summary

• Automatic brain tissue segmentation of neonatal
  MRI
• Detects white matter as myelinated and
  non-myelinated structures
• Makes use of prior knowledge
• Image intensity ordering
• Spatial locations (probabilistic atlas prior)
• To be used in two large UNC neonatal MRI studies
• Silvio Conte Center 125 neonates at risk
• Neonate Twin study (heritability)
• Current focus Validation

22
Acknowledgements

• Elizabeth Bullitt
• Petra Hüppi
• Koen van Leemput
• Insight Toolkit Community

Neoseg v1.0b
23
Validation (in progress)

• A) Semiautomated expert segmentation of a few
  cases
• Edge-based segmentation
• Level-set evolution
• Manual editing
• Primarily White-gray contour
• B) Simulated MRI data (similar to MNI ICBM)

24
New Probabilistic Atlas for the 2yrs group
14 subjects, aligned, intensity adjusted,
segmented (UNC M. Jomier/Piven/Cody/Gimpel/Gerig)