Multi-context Fuzzy Clustering for MRI Tissue Segmentation of Human Brain

1
Multi-context Fuzzy Clustering for MRI Tissue
Segmentation of Human Brain

• Chaozhe Zhu
• National Laboratory of Pattern Recognition, China

2
What is tissue segmentation?


Data information
Knowledge
Decision
3
Why segmentation?

• Quantification of tissue
  volume
• Visualization analysis
• Detection of pathology
• Monitoring lesion progression treatment
• Brain functional mapping
• Surgical navigation

3-D visualization of WM
Cerebral atrophy
Localising activation onto a high-resolution image
MS lesion
4
Intensity inhomogeneities in MR Images

• Artificial intensity variations (AIV)

Intra-tissue Intensity Variation (ITIV)
Appearance continuous, slowly varying shading
effect over the whole image domain Cause
hardware imperfections of MRI devices, such as
RF inhomogeneity
Appearance inherent intensity variation in the
brain, not artifact. Cause At different
positions in the brain, tissue attributes such
as T1, T2 time ,and densities are more or less
different .
5
http//pet.med.va.gov8080/papers/tech_reports/bia
s_correction/bias_1sub_fig8.html
6
(No Transcript)
7
(No Transcript)
8
Existing image models
Real intensities in the same tissue class are
identical no matter where are the pixels.
AIV

• Nothing of ITIV has been considered! (or
  partially modeled by measurement noise)
• Have to estimate the inhomogeniety!

9
A Novel Model
Previous model
Our model
Real intensities in the same tissue class can
vary with different position
AIV
ITIV
10
Basic assumptions
A1) Bias field (AIV) is smooth and slowly
varying. A2) Within a context, the classes of
tissues exist all together and there are
considerable pixels in each tissue class. A3)
Within a context, all pixels of the same tissue
have similar true intensities.
highly convoluted spatial distribution of
different tissues in brain
11
Model Simplification
Inhomogeneity is removed!
12
Clustering in single context
The assumptions are correct only in a sense of
statistics, accordingly, the simplified model is
also correct only in a sense of statistics.While
certain context, though just a few, may not
satisfy the assumptions because of
--Partial occupation. --Distribution
variation.
13
Our answer multi-context soft classification
Multi-context
Single Context
Statistical fusion
14
Experimental results (1)

• Single-context v.s.
  Multi-context

15
Experimental results (2)

• Intensity inhomogeneities (AIV)

FCM
MCFC
ORIGINAL
CORRUPTED
16
Experimental results (3)

• Intensity inhomogeneities (ITIV)

FCM
MCFC
ORIGINAL
17
Experimental results (4)
Fuzzy Mapping
GM
CSF
Hard result
WM
MCFC
FCM
18
Experimental results (5)
FCM
MCFC
TRUTH
Data from McConnell Brain Imaging Center at the
Montreal Neurological Institute
19
The relationship between MCR and context size
20
Computational time and Context size
21
Conclusion

• A new image model considering the intensity
  variation in real images.
• The estimation for intensity inhomogeneities is
  not indispensable any more.
• More compatible with human-visual perception.
• Only one parameter ( context size) to set.
• Easy to be implemented in a parallel way for real
  time applications

MCFC is a framework that can embed FCM as well as
any other soft-segmentation methods to calculate
the membership degree in each context and then
integrate all the soft decisions to make the
final decision.
22
Road ahead
Performance of MCFC

• Amount, Size, Shape of Contexts

23
Thanks! czzhu_at_nlpr.ia.ac.cn