Incorporating Non-rigid Registration into Expectation Maximization Algorithm to Segment MR Images
By
K.M. Pohl, W.M. Wells, A. Guimond, K. Kasai, M.E. Shenton, R.
Kikinis, W.E.L. Grimson, and S.K. Warfield
Email: [email protected]
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 2 -
Overview
• Introduction
• Incorporating Local Prior in EM-MF
• Current Implementation – Tools and Tricks
• Possible Advancements
• Conclusion
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 3 -
Goal
SPGR
T2W
Tissue Atlases
The Magic AutomaticSegmenter
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 4 -
EM-MF AlgorithmM-Step
E-Step
Smooth Bias
Image
Correct Intensities
MF-StepRegularize Weights
Estimate TissueProbability
Label Map
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 5 -
Mean Entropy Atlas
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 6 -
Merging MEA with SPGR
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 7 -
Bias
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 8 -
Bias in Color
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 9 -
3 D View of SPGR
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 10 -
Including Local Priors
2. Step
1. Step
Bra
in A
tlas
Cas
e
Registration
Pro
bab
ility
Map
s
Align Atlas
3. Step
M-Step
E-Step
Bias
Correct
MF-Step
Estimate
Label Map
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 11 -
eP(Tissue T)
* P(GV[x][y][z] | Distribution of T,Bias)
EM Algorithm
P(Tissue T | Position [x][y][z])
Local Prior
Estimating the Tissue Class
WT[x][y][z]
* eEnergy(WT[x][y][z] | Neighboring W)
MF-Approximation
GV[x][y][z] = Grey Value at position [x][y][z] WT [x][y][z] = Weights for tissue class T at position [x][y][z]
+
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 12 -
Comparing different Segmenter
Registration
only
EM
-MF
Affin
e EM
-MF
Non
Rig
id E
M-M
F
2 Channel Input - Segmenting up to 7 tissue classes
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 13 -
2 Channel Segmentation with Patient Case and 11 Tissue Classes
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 14 -
Correction of 1 Channel EM-MF-LP through Specialist
Background CSVSkin Grey Matter White Matter
Right/Left Amygdala Right/Left Superior Temporalgyrus
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 15 -
Comparing Manual to EM-MF-LP of the STG
Rater A Rater BE
M-M
F-L
P
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 16 -
Current Installation• Algorithm is a VTK Filter integrated in Slicer • MF Approximation:
– Multi Threaded
– Lookup Table for Gaussian Distribution
• Using several Relaxation Methods instead of the Mean Field Energy Function
• Multi Channel Input (SPGR, T2 , PD)• Train Tissue Definition, e.g. CIM, Distribution • Interface to Matlab
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 17 -
EM-MF in Slicer
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 18 -
Tabs of GUI Overview Class Definition Class Interaction EM Settings
Skill Level
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 19 -
Possible Improvements• Registration Step:
– After each segmentation re-register case with atlas
• E Step– Include shape and topology information in weight
calculation
– Use local class interaction matrix
• M Step:– Use several other filters to smooth bias, e.g. Box Filter,
Pascal Triangle, …
– Include “trash tissue class” where pixels get assigned if all weights are low Bias does not get corrupted
MICCAI’02 Incorporating Non-rigid Registration into Expectation
Maximization Algorithm to Segment MR Images- 20 -
Conclusion
• Made EM-MF Algorithm more robust• Segmented tissue classes with overlapping gray
value distributions• Included spatial/atlas information into E-Step• Cortex pacellation possible• Future Work: Validating Segmentation