A Unified Feature Registration Framework for Brain Anatomical Alignment

1
A Unified Feature RegistrationFramework for
Brain Anatomical Alignment

• Haili Chui, Robert Schultz, Lawrence Win, James
  Duncan and Anand Rangarajan

Image Processing and Analysis Group Departments
of Electrical Engineering and Diagnostic
Radiology Yale University Department of
Computer Information Science and
Engineering University of Florida
2
Brain Anatomical Alignment

• Brains are different
• Shape.
• Structure.
• Direct comparison of brains between different
  subjects is not very accurate.
• Statistically and quantitatively more accurate
  study requires the brain image data to be put in
  a common normalized space through alignment.
• Examples of areas that need brain registration
• Studying structure-function connection.
• Tracking temporal changes.
• Generating probabilistic atlases.
• Creating deformable atlases.

3
Studying Function-Structure Connection
4
Inter-Subject Brain Registration

• Inter-subject brain registration
• Alignment of brain MRI images from different
  subjects to remove some of the shape variability.
• Difficulties
• Complexity of the brain structure.
• Variability between brains.

• Brain feature registration
• Choose a few salient structural features as a
  concise representation of the brain for matching.
• Overcome complexity only model important
  structural features.
• Overcome variability only model consistent
  features.

5
Previous Work 3D Sulcal Point Matching
Feature Extraction
Extracted Point Features
6
Previous Work 3D Sulcal Point Matching
Overlay of 5 subjects before TPS alignment
7
A Unified Feature Registration Method
8
Non-rigid Feature Point Registration
9
Unification of Different Features

• Ability to incorporate different types of
  geometrical features.
• Points.
• Curves.
• Open surface ribbons.
• Closed surfaces.
• Simultaneously register all features --- utilize
  the spatial inter-relationship between different
  features to improve registration.

10
Joint Clustering-Matching Algorithm (JCM)
11
Overcome Sub-sampling Problem

• Sub-sampling (e.g. clustering) reduces
  computational cost for matching.

• In-consistency problem with sub-sampling

• The in-consistency can be overcome by
  sub-sampling (clustering) and matching
  simultaneously.

12
Joint Clustering-Matching Algorithm (JCM)

• Diagram

• JCM
• Reduce computational cost using sub-sampled
  cluster centers.
• Accomplish optimal cluster placement through
  joint clustering and matching.
• Symmetric two way matching.

13
JCM Energy Function
Point Set X
Point Set Y
Clustering
Clustering
Clusters Center Set V
Cluster Center Set U
Annealing
14
JCM Energy Function

• Clustering and regularization energy function

• First two terms perform clustering, next four
  perform non-rigid matching and last two are
  entropy terms.

15
JCM Example

• Matching 2 face patterns with JCM (click to play
  movie).

16
Experiments
17
Comparison of Different Features

• Different features can be used in our approach.

18
Synthetic Study Setup
19
Results Method I vs. Method III

• Outer cortical surface alone can not provide
  adequate information for sub-cortical structures.
  
• Combination of two features works better.

20
Results Method II vs. Method III

• Major sulcal ribbons alone are too sparse --- the
  brain structures that are relatively far away
  from the ribbons got poorly aligned.
• Combination of two features works better.

21
Conclusion

• Combination of different features improves
  registration.
• Unified brain feature registration approach
• Capable of estimating non-rigid transformations
  without the correspondence information.
• General unified framework.
• Symmetric.
• Efficient.

22
Acknowledgements

• Members of the Image Processing and Analysis
  Group at Yale University
• Hemant Tagare.
• Lawrence Staib.
• Xiaolan Zeng.
• Xenios Papademetris.
• Oskar Skrinjar.
• Yongmei Wang.
• Colleagues in the brain registration project
• Joseph Walline.
• Partially supported is by grants from the
  Whitaker Foundation, NSF, and NIH.

23
Future Work
24
Estimating An Average Shape

• Given multiple sample shapes (sample point sets),
  compute the average shape for which the joint
  distance between the samples and the average is
  the shortest.

Average ?

• Difficult if the correspondences between the
  sample points are unknown.

25
Super Clustering-Matching Algorithm (SCM)

• Diagram

26
(No Transcript)
27
End

• Further Information
• Web site http//noodle.med.yale.edu/chui/

28
End
29
2D Examples of RPM
30
(No Transcript)
31
(No Transcript)
32
Example Application Face Matching
33
Example Application Face Matching