1. Advisor: Chin-Chen Chang1, 2
The study of segmentation and hiding techniques for medical images 醫學影像之切割與隱藏技術之研究
Advisor: Chin-Chen Chang1, 2
Student: Pei-Yan Pai2
1 Dept. of Information Engineering and Computer Science, Feng Chia University
2 Dept. of Computer Science, National Tsing Hua University

2. Outline Part I: Image segmentation Part II: Image hiding
Adaptable threshold detector (ATD)
Part II: Image hiding
An ROI-based medical image hiding method

3. Part I: Image segmentation
Adaptable threshold detector (ATD)

4. Introduction Image segmentation Thresholding method
Segment interested object
Medical image processing, pattern recognition, etc.
Thresholding method
Simple and effective
Otsu’s thresholding method (OTM)
Threshold=156
Cervical smear image
Gray-level histogram
Contoured image
Binary image

5. Motivation Drawbacks of Otsu’s thresholding method
The failed threshold
The cluster with a larger variance
A larger quantity of data.
Threshold is 153 by using OTM
Optimal threshold is 156
Cervical smear image
Gray-level histogram
The contour obtained by using OTM

6. Motivation (cont.) Drawbacks of Otsu’s thresholding method (cont.)
The fixed threshold
Original image
Cytoplast
Nucleus
Result by using OTM

7. Otsu’s thresholding method (OTM)
Minimizes within-class variance
L=255
P: Probability
σ2: Variance
Select threshold by using OTM
Optimal threshold
Select threshold by using OTM
Optimal threshold C1 C2 C1 C2
255
Gray-level
Gray-level
255

8. The principle of adaptable threshold detector (ATD)
SD: Standard deviation
R: Group interval C1 C1 C2 50
170
210
255
RC1
RC2
170-50=120
=40
Within-class standard deviation (WCSD)
One threshold
Multi-thresholds

9. Genetic-based parameter detector (GBPD) (1/2)
A chromosome
Fitness function
Takes average “misclassification error” (ME), “modified Hausdorff distance” (MHD), or “relative foreground area error” (RAE) as the measure of fitness
Selection operator
Randomly select N optimal chromosomes from reserved chromosomes
N and 2×N’ chromosomes generated in the mutation and crossover operations
|s1|=|s2|=14, gap1 = gap2 = 0.1,
r1 = 0.1×4 = 0.4
r2 = 0.1×6 = 0.6
ni: The number of 1-bits in ith substring
gapi: The maximal estimate errors of ri

10. Genetic-based parameter detector (GBPD) (2/2)
Mutation operator
Crossover operator

11. Experimental results (1/18)
The parameters in GBPD : |s1|=|s2|=40, N=N´=40, and gap1=gap2=0.1
Compared methods:
Otsu’s threholding method1 (OTM)
Valley-emphasis method2 (VEM)
Minimum class variance thresholding method3 (MCVTM)
1Otsu, N., “A threshold selection method from gray-level histograms,” IEEE Transactions of Systems, Man, and Cybernetics, vol. 9, no. 1, pp.62-66, 1979.
2Ng, H. F., “Automatic thresholding for defect detection,” Pattern Recognition Letters, vol. 27, no. 14, pp , 2006.
3Hou, Z., Hu, Q., and Nowinski, W. L., “On minimum variance thresholding,” Pattern Recognition letters, vol. 27, no. 14, pp , 2006.

12. Experimental results (2/18)
Four 182×182 test images
Image 1
Gray-level histogram of Image 1
Image 2
Gray-level histogram of Image 2
Gray-level histogram of Image 4
Image 3
Gray-level histogram of Image 3
Image 4

13. Experimental results (3/18)
r1 = 2.2 and r2 = 2.0, train images: 1, 4

14. Experimental results (4/18)
(a) ME
(b) RAE
(c) MHD

15. Experimental results (5/18)
Three 128×128 test images
Image 1
Gray-level histogram of Image 1
Image 2
Gray-level histogram of Image 2
Image 3
Gray-level histogram of Image 3

16. Experimental results (6/18)
r1 = 2.2 and r2 = 0, train images: 1, 2

17. Experimental results (7/18)
(a) ME
(b) RAE
(c) MHD

18. Experimental results (8/18)
Eight 128×128 test images (cervical smear images)
Image 1
Gray-level histogram of Image 1
Image 2
Gray-level histogram of Image 2
Image 3
Gray-level histogram of Image 3
Image 4
Gray-level histogram of Image 4
Image 5
Gray-level histogram of Image 5
Image 6
Gray-level histogram of Image 6
Image 7
Gray-level histogram of Image 7
Image 8
Gray-level histogram of Image 8

19. Experimental results (9/18)
r1 = 2.0 and r2 = 0.8, train images: 1, 2, 5, 6

20. Experimental results (10/18)
(a) ME
(b) RAE
(c) MHD

21. Experimental results (11/18)
Two 128×128 cervical smear images add 1%, 1.5%, and 2% of Gaussian noises (GN).
Image 1(GN=1%)
Gray-level histogram of Image 1(GN=1%)
Image 1(GN=1.5%)
Gray-level histogram of Image 1(GN=1.5%)
Image 1(GN=2%)
Gray-level histogram of Image 1(GN=2%)
Image 7(GN=1%)
Gray-level histogram of Image 7(GN=1%)
Image 7(GN=1.5%)
Gray-level histogram of Image 7(GN=1.5%)
Image 7(GN=2%)
Gray-level histogram of Image 7(GN=2%)

22. Experimental results (12/18)
r1 = 2.0 and r2 = 0.8

23. Experimental results (13/18)
(a) PSNR
(b) ME
(c) RAE
(d) MHD

24. Experimental results (14/18)
Eight 128×128 test images (cervical smear images)

25. Experimental results (15/18)
r1 = 2.9 and r2 = 1.7,
train images: 1, 2, 5, 6

26. Experimental results (16/18)
(a) ME
(b) RAE
Average segmentation errors using ME, RAE, and MHD as fitness
Method ME
RAE
MHD
ATD using ME
ATD using RAE
ATD using MHD
OTM
VEM
MCVTM
(c) MHD

27. Experimental results (17/18)
Two 256×256 pixel blood smear images
Image 1
Gray-level histogram of Image 1
Image 2
Gray-level histogram of Image 2
r1 = 2.9 and r2 = 1.7, train image: 1

28. Experimental results (18/18)
(a) ME
(b) RAE
(c) MHD

29. Part II: Image hiding
An ROI-based medical image hiding method

30. Magnetic resonance imaging (MRI) brain image
Introduction
Image hiding
Embedding phase
Embedding
Magnetic resonance imaging (MRI) brain image
Stego-image
Secret data
0101..
Extracting phase
Irreversible image
hiding
Secret data
0101..
Extracting
Reversible image
hiding
Stego-image
MRI brain image

31. Motivation ROI Irreversible V.S. reversible image hiding methods
High embedding capacity
Cannot competently restore the cover image
Reversible
Limited embedding capacity
The cover image can be recovered without any distortion
Medical image properties
ROI (Region of interesting)
Background is unimportant
Goals
Automatically segment ROI
Increase embedding capacity
ROI
MRI brain image

32. Flowchart of proposed method
Secret data
0101..
Embedding
ROI
Reversible image
hiding
ATD
Irreversible image
hiding
Non-ROI
MRI brain image
ROI segmented image
Stego-image

33. The proposed method (1/9)
ROI segmentation
ATD T*
Record the coordinates as secret data
MRI brain image
Its gray-level histogram
Binary image
ROI image

34. The proposed method (2/9) -Embedding phase
Non-ROI
Simple LSB substitution (SLSB)
Simple
Good stego-image quality 8 15 3 10 5 7 1
255
154
156
155
152
151
153
146 4 2
150
178 6
Secret data ….
k=3 7 1 2
Case 2
Case 1
Case 3

35. The proposed method (3/9) -Embedding phase
ROI
High payload frequency-based reversible image hiding (HPFRIH) method LL HL
255
154
156
155
152
151
153
146
150
178
255
153.5
154
151.25 -1
0.5
1.25
152.75
153.25
0.75
0.25
154.25
151.75
-0.75
-1.25
153
158.25
-6.5
1.5
-0.5
1.75 1
-2.25
-0.25
-1.5
-6.75 7
One-level
HDWT
ROI LH HH

36. The proposed method (4/9) -Embedding phase
Rules:
1. I(x, y) < 2k′ then
I′ (x, y)=I(x, y)+2k′ and BM[x, y] = 1
2. I(x, y) >255-2k′ then
I′ (x, y)=I(x, y)-2k′ and BM[x, y] = 1
3. Otherwise
I′ (x, y)=I(x, y) and BM[x, y] = 0
Pre-process
Solve underflow/overflow problem
2k′
255-2k′
k′=3
255
154
156
155
152
151
153
146
150
178
247
154
156
155
152
151
153
146
150
178 1
ROI′
ROI BM

37. The proposed method (5/9) -Embedding phase
HDWT LL HL
247
154
156
155
152
151
153
146
150
178
247
153.5
154
151.25 -1
0.5
1.25
152.75
153.25
0.75
0.25
154.25
151.75
-0.75
-1.25
153
158.25
-6.5
1.5
-0.5
1.75 1
-2.25
-0.25
-1.5
-6.75 7
One-level
HDWT
ROI LH HH

38. The proposed method (6/9) -Embedding phase
247
153.5
154
151.25 -1
0.5
1.25
152.75
153.25
0.75
0.25
154.25
151.75
-0.75
-1.25
153
158.25
-6.5
1.5
-0.5
1.75 1
-2.25
-0.25
-1.5
-6.75 7
247
153.5
154
151.25 -1
0.5
1.25
152.75
153.25
0.75
0.25
154.25
151.75
-0.75
-1.25
153
158.25
-6.5
1.5
-0.5
1.75 1
-2.25
-0.25
-1.5
-6.75 7
IHDWT
ROI′′
HHS
HHI′
HHD
HHS
HHI
HHD 1 1 2 7
0.25
0.75
0.5
k′-LSBs of matrix HHI
k′=3 0: 1:
ED={ } + BM
Bits to be hidden: 001
LSBs replacement
Adaptive arithmetic
coding
Compressed data +
Secret bits

39. The proposed method (7/9) -Extracting phase
Non-ROI extracting phase
Take k-LSBs of a pixel in non-ROI as secret data
EX:
k=3
I′ (x, y)=7=
Secret data

40. The proposed method (8/9) -Extracting phase
ROI extracting phase
k′=3
247
153.5
154
151.25 -1
0.5
1.25
152.75
153.25
0.75
0.25
154.25
151.75
-0.75
-1.25
153
158.25
-6.5
1.5
-0.5
1.75 1
-2.25
-0.25
-1.5
-6.75 7
HDWT
247
153.5
154
151.25 -1
0.5
1.25
152.75
153.25
0.75
0.25
154.25
151.75
-0.75
-1.25
153
158.25
-6.5
1.5
-0.5
1.75 1
-2.25
-0.25
-1.5
-6.75 7
IHDWT
ROI′′
ROI′
HHS
HHI′
HHD
HHS
HHI
HHD 1 1 2 7
0.25
0.75
0.5
k′-LSBs of matrix HHI′
ED={ } BM
Adaptive arithmetic
decoding
Compressed data +
Secret bits

41. The proposed method (9/9) -Extracting phase
Rules:
IF I′(x, y) < 22k′ and BM[x, y] = 1
I(x, y)=I′(x, y)-2k′ and BM[x, y] = 1
Else if 255-22k′<I′(x, y) ≤ 255 then
I′(x, y)=I′(x, y)+2k′ and BM[x, y] = 1
Else
I′(x, y)=I′(x, y) and BM[x, y] = 0
ROI extracting phase (cont.)
k′=3
247
154
156
155
152
151
153
146
150
178 1
255
154
156
155
152
151
153
146
150
178
ROI′ BM
ROI

42. Experimental results (1/5)
The performances of HPFRIH method
The stego-images of “Lena” and “Baboon”
Image size: 512 ×512
k′=3
(a) Lena (PSNR=36.24)
(b) Baboon (PSNR=36.33)

43. Experimental results (2/5)
The performances of HPFRIH method (cont.)
Table 1. The performance comparison among the different reversible image hiding methods
PHC: Pure hiding capacity

44. Experimental results (3/5)
The performances of proposed method
Image size: 256 ×256
The test images

45. Experimental results (4/5)
The performances of proposed method (cont.)
r1 = 2.0 and r2 = 0.8
r1 = 2.0 and r2 = 1.7
The stego-images (k=k′=3)

46. Experimental results (5/5)
The performances of proposed method (cont.)
The PSNRs and Rs of ROI-based and other image hiding methods

47. Conclusions and future works
The adaptable threshold detector (ATD)
Overcome the drawbacks of OTM
Good performance
Adaptation
An ROI-based medical image hiding method
High embedding capacity
ROI can be completely restored
Future works
Image segmentation
Cell counting
Cell segmentation
Image hiding for medical image
Increase embedding capacity
Apoptosis
Fluorescence microscopic image

48. Thanks for your attention...