Robust Segmentation of Brain MRI using Combination of Registration and EMbased methods

1
Robust Segmentation of Brain MRI using
Combination of Registration- and EM-based methods

• Maria Murgasova1, Leigh Dyet2,
• David Edwards2, Mary Rutherford2, Jo Hajnal2 and
  Daniel Rueckert1
• 1Department of Computing
• 2Department of Imaging Sciences
• Imperial College London

2
Outline

• Medical background
• Registration-based segmentation
• EM-based segmentation
• Building a population-specific probabilistic
  atlas
• Building a subject-specific probabilistic atlas
• Comparison of the methods

3
Motivation

• The effect of premature birth
• impaired brain development
• neurological, behavioural, learning difficulties
• To understand and treat the changes we need to
  measure
• volumes of different brain structures
• growth of different brain structures
• This requires
• segmentation of anatomical
• structures at different time points

Brain MRI of a 2-year-old child
4
Non-rigid registration

• The aim to warp one image to another
• 3D B-spline-based registration (Rueckert 1999)

Reference subject
New subject
Affine registration
B-spline registration
5
Registration-based segmentation

• Non-rigid registration of atlas to subject
• Segmentation is warped from atlas to subject
• Advantage
• Does not assume any tissue intensity model
• gt successful in central brain structures
• Disadvantage
• Does not always deal well with complex cortical
  folding

6
EM-based segmentation

• Classical model for brain MRI
• 3 basic tissue classes (WM, GM, CSF)
• Tissue intensity distributions approximately
  Gaussian
• Advantage
• Can capture complexity of cortex based on
  intensity
• Disadvantage
• Cant deal well with overlaps in tissue intensity
  distributions

Real tissue distributions of 2-year-old subject
based on manual segmentation
7
EM-based segmentation

• Sub cortical structures brighter than cortical
• Overlaps cause significant difficulties
• Classical WM, GM, CSF model not sufficient for
  correct segmentation

8
EM-based segmentation

• Probabilistic atlas
• Aligned with the image
• Spatially constrains the segmentation process
• Helps to overcome misclassification due to
  overlaps in tissue intensity distributions

9
EM-based segmentation

• Affine registration of probabilistic atlas
• EM segmentation (Van Leemput 1999)
• E-step soft segmentation
• M-step Gaussian tissue intensity distribution

10
Application of EM to young children

• Shape of brain in young children different from
  adults
• White matter overestimated in sub-cortical area
• Requires a specific atlas for young children

Segmentation of 1-year-old
Adult
Adult atlas
1-year-old
11
Creating a population-specific atlas

• Start with 1 manual segmentation

Training images
Reference subject
Probabilistic atlas
Reference subject
Affine registration
Non-rigid registration
Average segmentations
12
Segmentation results

• Improvement in thalamus
• 1.0T brain MRI of a 2-year-old child

Image
Manual segmentation
EM with the adult atlas
EM with the new atlas
13
Limitations

• Registration of atlas to image
• Affine not always flexible enough
• Non-rigid does not work

Brain with a very different shape (focal lesions)
EM segmentation into 11 structures
14
Creating a subject-specific atlas

• Start with 1 manual segmentation

Training images
New subject
Probabilistic atlas
Reference subject
Non-rigid registration
Non-rigid registration
Average segmentations
15
Robust registration-based segmentation

• Subject-specific atlas can be used as a final
  segmentation
• Improvement over registration-based segmentation
• Some loss of detail

Registration-based segmentation
Robust registration-based segmentation
16
Robust EM-based segmentation

• Subject-specific atlas as prior information
• Correct structure boundaries
• More detail included in segmentation

EM-based segmentation
Robust EM-based segmentation
17
Robust EM-based segmentation

• Successful on brains with significantly different
  shape
• Example focal lesions

Robust EM-based segmentation into 11 structures
Failed EM-based segmentation into 3 structures
18
Validation

• 4 subjects 5 slices
• Dice metrics

• Agreement between manual and automatic
  segmentation
• Tman set of samples in manual segmentation
• Taut set of samples in automatic segmentation

19
Acknowledgements

• This work is funded by

20
(No Transcript)