Medical imaging applications produce large sets of similar images. The huge amount of data makes the manual analysis and interpretation a fastidious task. Medical image segmentation is thus an important process in image processing used to partition the images into different regions (e.g. gray matter, white matter and cerebrospinal fluid). Hidden Markov Random Field (HMRF) Model and Gibbs distributions provide powerful tools for image modeling. In this paper, we use a HMRF model to perform segmentation of volumetric medical images. We have a problem with incomplete data. We seek the segmented images according to the MAP (Maximum A Posteriori) criterion. MAP estimation leads to the minimization of an energy function. This problem is computationally intractable. Therefore, optimizations techniques are used to compute a solution. We will evaluate the segmentation upon two major factors: the time of calculation and the quality of segmentation. Processing time is reduced by distributing the computation of segmentation on a powerful and inexpensive architecture that consists of a cluster of personal computers. Parallel programming was done by using the standard MPI (Message Passing Interface).

1. The Third International Conference on Digital Information Processing and
Communications (ICDIPC 2013)
Medical Image Segmentation Using Hidden Markov Random Field
A Distributed Approach
Theme

2.  INTRODUCTION
 SEGMENTATION BY USING HMRF
 EXPERIMENTAL RESULTS
 CONCLUSION AND PERSPECTIVES
2

3. One exam by CT (Computed Tomography)
scanner can produce hundred images.
All of these images
represents a 3D
image
Processing and analysis of these images
becomes a difficult and daunting task
The classical analysis of medical cuts
3
I N T R O D U C T I O N
Problem

4. 3D automatic
segmentation
The 3D image The segmented 3D image
4
I N T R O D U C T I O N
Solution :
Tool to aid the physician to make the decision
based on Automatic segmentation.

5. 5
T H E A I M
Relevance of the physician aid tool to
make the decision based on
The time of computation The quality of segmentation
TIME + QUALITY

6.  INTRODUCTION
 SEGMENTATION BY USING HMRF
 EXPERIMENTAL RESULTS
 CONCLUSION AND PERSPECTIVES
6

7. S E G M E N TAT I O N B Y U S I N G H M R F
7
1 2
3 4
Y: Observed Image
X: Hidden Image
 

2C,s
2
s
),(2-(1)2ln(
2
)²-(y
y)(x,
t
tsx
Ss x
x
xx
Ts
s
s





 y)(x,minarg 
Xx
x


8. S E G M E N TAT I O N B Y U S I N G H M R F
8
Optimizations techniques are used like ICM, …
 Problem
Minimizing the function (x,y) is computationally intractable.
 Solution

9. S E G M E N TAT I O N B Y U S I N G H M R F
ICM Algorithm:
1. Initialization: Start with an arbitrary labeling x0 and let n=0.
2. At step n:
Visit all the sites according to a visiting scheme and in every site :
,
3. Increment n. Goto 2, until a stopping criterion is satisfied.
9
( )
1
arg min ( )card S
n
s s s
x
x U x 

   

10.  INTRODUCTION
 SEGMENTATION BY USING HMRF
 EXPERIMENTAL RESULTS
 CONCLUSION AND PERSPECTIVES
10

11. E X P E R I M E N TA L R E S U LT S
11
Configuration Hardware :
The cluster of eight identical machines
 Switch (Catalyst 3560G)
Configuration Software:
The Parallelization library is Open MPI
Platform application framework Qt
Linux system (ubuntu 11.04)

12. E X P E R I M E N TA L R E S U LT S
12
Benchmark Name of benchmark Dimension Link
1
MRI Phantom 8Bits
(t1_icbm_normal_1mm_pn0
_rf0.rawb)
181 x 217 x 181
http://mouldy.bic.mni.
mcgill.ca/brainweb/anat
omic_normal.html
2
Head MRT Angiography
8Bits
(mrt8_angio2.raw)
256 x 320 x 128
http://www.gris.uni-
tuebingen.de/edu/areas/s
civis/volren/datasets/ne
w.html
3 Head MRI CISS 8Bits
(mri_ventricles.raw) 256 x 256 x 124
http://www.gris.uni-
tuebingen.de/edu/areas/s
civis/volren/datasets/ne
w.html
Benchmarks images used in our tests.

13. E X P E R I M E N TA L R E S U LT S
13
Visual results
Benchmark : 1

14. E X P E R I M E N TA L R E S U LT S
14
Visual results
Benchmark : 2

15. E X P E R I M E N TA L R E S U LT S
15
Visual results
Benchmark : 3

16. Evaluating the quality of the segmentation
16
FNFPTP2
TP2DC


Kappa index
Ground
truth
The image to
segment
The segmented
image
E X P E R I M E N TA L R E S U LT S

17. E X P E R I M E N TA L R E S U LT S
Comparison : Mean kappa index values
Benchmark : 1
Slices : 90-119
Methods : Otsu, MoG, MoGG and our method
17
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
White
Matter
Gray
Matter
CSF Matter
Otsu
MoG
MoGG
Our Method
Methods
Kappa Index

18. Speed-up
18
E X P E R I M E N TA L R E S U LT S
Processing Time

19. 19
E X P E R I M E N TA L R E S U LT S
Processing Time
0
1
2
3
4
5
6
7
8
9
1 PC 2 PCs 4 PCs 8 PCs
Benchmark 1
Benchmark 2
Benchmark 3
Time (h)
Number of PCs
Benchmarks

20. 20
E X P E R I M E N TA L R E S U LT S
SPEED UP
0
1
2
3
4
5
6
7
8
9
1 PC 2 PCs 4 PCs 8 PCs
Benchmark 1
Benchmark 2
Benchmark 3
Speed-up
Number of PCs
Benchmarks

21.  INTRODUCTION
 SEGMENTATION BY USING HMRF
 EXPERIMENTAL RESULTS
 CONCLUSION AND PERSPECTIVES
21

22.  The kappa index can be used only when we know beforehand
segmentation ground truth .
 In our tests we notice our implemented method seems generally
better than the thresholding-based segmentation methods (Otsu, MoG,
MoGG ).
 The processing time is improved by the use of a cluster of PCs.
22
C O N C L U S I O N A N D P E R S P E C T I V E S

23. However, further work must take into account like :
 The cluster of PCs must be incremented to see the limits of its
contribution.
 Comparison with other methods
 Implementation of other optimization methods
23
C O N C L U S I O N A N D P E R S P E C T I V E S

24. Thank you for your
attention