This document provides an overview of tissue segmentation methods using 2D histogram matching in a sequence of MR brain images. It discusses:
1. The main idea is to apply 2D histogram matching to calculate training matrices for each tissue type in a test MR image, and then classify each test 2D histogram using the training matrices.
2. Challenges include tissue intensity variations between patients/scanners and changes along image sequences due to artifacts.
3. The method contributes algorithms for MR image sequence division, 2D histogram matching using vector specification, and image segmentation via back projection of classified histograms.
Tissue Segmentation Methods Using 2D Histogram Matching in a Sequence of MR Brain Images_part1
1. Tissue Segmentation Methods using
2D Histogram Matching in a Sequence
of MR Brain Images (Part 1)
Vladimir Kanchev, PhD
Radiocommunications and
Videotechnologies Department
TU Sofia, Sofia, Bulgaria
July 2017
2. Page 2
This Research is Reported in:
Kanchev, Vladimir and
Roumen Kountchev.
"Tissue Segmentation Methods Using
2D Histogram Matching in a Sequence
of MR Brain Images."
New Approaches in Intelligent Image
Analysis. Springer International
Publishing, 2016. 183-222.
(Chapter 6)
3. Page 3
Contents
1. Main idea and contributions
2. Introduction
3. Method description
4. Experimental results
5. Conclusions and future work
4. Page 4
Main Objective
The volume of brain tissues depends on the age,
health condition and sex of the patient. Additionally,
some diseases and health conditions lead to a
change and a decrease in the tissue volumes.
Therefore, their precise measurement is important
for the patient diagnosis.
5. Page 5
MR Image Segmentation
Def: Image segmentation aims to divide an
image into non-overlapping, constituent
regions, homogeneous to some feature as
intensity and texture*. Classification aims
to assign to each image pixel a tissue class
or air. Both operations are interconnected.
Purpose: Assist radiologists in the detection of brain
diseases based on MR images – provide “a
second opinion”.
* Pham et al. "Current methods in medical image segmentation”. Annual
review of biomedical engineering 2.1 (2000): 315-337.
6. Page 6
MR Brain Image Segmentation
* Goossens, B., & Philips, W. (2015). MRI segmentation of the human
brain: challenges, methods, and applications. Computational and
mathematical methods in medicine, 2015.
*
7. Page 7
Main Idea
In the chapter we suggest an MR image
segmentation method using a 2D histogram
matching.
We apply the 2D histogram matching to
(1) calculate train matrices for each tissue of the
test MR image. Then we (2) classify a test
2D histogram of each test MR image using them.
8. Page 8
Other Challenges
Tissue intensity distributions vary depending on
the patient, scanner and scanning conditions
Tissue properties (area, texture, intensity
distribution, etc.) change along an MR image
sequence
Artefacts modify the intensity distributions of
brain tissues upon few neighboring MR images
9. Page 9
Main Contributions
Our method includes algorithms for:
MR image sequence division using a normalized
wave hedges distance
2D histogram matching using a vector
specification
MR image segmentation using a back projection
of a classified 2D histogram
10. Page 10
Contents
1. Main idea and contributions
2. Introduction
3. Method description
4. Experimental results
5. Conclusions and future work
12. Page 12
MRI Data Properties
*
* http://images.wisegeek.com/mri-machine-with-screens.jpg
13. Page 13
MRI Data Properties
MR imaging is a leading medical technology for
visualization of internal human organs and
tissues
The intensity levels of tissues in MR images
depend on the quantity of hydrogen atoms
Presence of artefacts – bias field, chemical shift,
partial volume, ringing, etc.
Ever increasing quantity of available MRI data –
introduction of data-driven methods
14. Page 14
Different tissues in MR brain images
Main tissues:
CSF – CerebroSpinal Fluid
GM – Gray Matter
WM – White Matter
Other (redundant) tissues:
sculp, skull, dura, muscle, etc.
15. Page 15
Different tissues in MR brain images
*
*Ramon, Ceon et al. "Influence of head models on EEG simulations and inverse
source localizations." BioMedical Engineering Online 5.1 (2006): 10.
16. Page 16
MRI Data Properties
MR imaging (T1 - weighted) provides very good
resolution on soft brain tissues (CSF - the lowest
intensity values, GM - middle, WM - the highest)
Neighboring pixel pairs within tissues have equal
or close intensity values (piecewise smooth or
constant, slowly varying values)
Neighboring pixel pairs on the border between
tissues have far different intensity values
(piecewise discontinuous)
17. Page 17
MRI Data Properties
Variability of intensity values of separate tissues
within a patient/MRI scanning (intra-scan) or
between patients/MRI scanning (inter-scan)
Intra-scan variability necessitates intensity
standardization of MRI data
MR image noises and artefacts require filtering and
correction
18. Page 18
MR Image Sequence
*
* Anum Masood ‘Detection and Classification of Tumor in Magnetic
Resonance (MRI) Brain Images’ [Retrieved 13.07.2017]
19. Page 19
Bias Field
Bias field or Intensity inhomogeneity (INH):
represents a low-frequency and a very smooth
signal which distorts the MR image (in our case)
alters tissue intensity distribution in one or more
neighbouring MR images
blurs the MR image and reduces high frequency
content of the MR image as edges and contours
20. Page 20
Bias Field
*
* Vovk, Uros et al. “A Review of Methods for Correction of Intensity
Inhomogeneity in MRI.” IEEE transactions on medical imaging 26.3
(2007): 405-21.
21. Page 21
Partial Volume Effect
Makes border pixels/voxels with intensity
values of two neighboring tissues, which
depend on the proportions of each tissue
Occurs due to the limited resolution of an MR
image and image sampling
Causes blurring of the image and spill-out effect
22. Page 22
Partial Volume Effect
*
* Tohka Jussi: “Partial volume effect modeling for segmentation and tissue
classification of brain magnetic resonance images: A review”. World Journal of
Radiology 6.11 (2014):855
23. Page 23
MRI Artefacts
Bias field leads to:
a change in the intensity distribution of a single tissue
along the whole MR image
a change in the intensity range of the MR image
introduction of low-frequency noise in the MR image
Partial volume effect leads to:
blurring of neighboring tissue edges
the introduction of new intensity values and a change in
high frequencies
the introduction of additional transitive tissue classes
CSF-GM, CSF-WM, GM-WM into segmentation methods
24. Page 24
Properties
Presence of both artefacts:
breaks the assumption of piecewise intra-tissue
smoothness and inter-tissue discontinuity of
MRI data
leads to worse results of the segmentation
algorithms
is not easily recognizable from humans
26. Page 26
Transductive Learning
“When solving a problem of interest, do not solve a
more general problem as an intermediate step. Try
to get the answer that you really need but not a
more general one.”
Vladimir Vapnik (author of the idea for Transductive learning, SVM
classifier, Vapnik-Chervonenkis theory, etc.)
27. Page 27
Transductive Learning
Depending on whether we build a general model
for the classification of a test set of instances, we use:
inductive learning (build a model)
transductive learning (do not build a model)
28. Page 28
Transductive Learning
We have a*:
Training set of labeled examples
where is an input vector and –
a corresponding output label
Test set of unlabeled examples
We look for:
Expected labels for all instances in
),x( i i
yU
ix i
y
V
)x( '
i
'
i
y
V
* https://codesachin.wordpress.com/2016/07/03/a-small-and-easy-
introduction-to-transductive-learning/
29. Page 29
Transductive Learning
We build labels only for instances from :
Input for supervised learning (and , if we
go for semi-supervised learning).
Output The set of expected labels for all
instances in .
Approach We leverage the information contained
in the instances of (with respect to
those in ) to make better prediction
for specifically.
U
V
V
V
U
V
V
30. Page 30
Inductive Learning
We build a general model to classify unseen
instances in the future besides those in :
Input Train set for supervised learning (and
if we go for semi-supervised learning).
Output The set of expected labels for all
instances in test set .
Approach Compute a function , using information
in such that is as close to as
possible over all instances in . Using
the function, compute each as much
as .
U
V
V
f
U
)x( if i
y
'
i
y
)x( '
i
'
fyi
V
U
31. Page 31
Properties
Transductive learning:
is slower but more precise method than the
inductive learning method
cannot label new instances except those in
requires integrity of test set instances – no
additional noise
V
32. Page 32
Transductive Learning
How it relates to our segmentation method:
we build an ad-hoc model (model 2D histograms)
for each MR image subsequence
we adapt the model for each test 2D histogram,
using matching, thus we find out the train matrices
of each test 2D histogram (partial classification)
34. Page 34
Histogram Matching
Histogram matching or specification *:
Def: A method used to generate a processed
image that has a specified histogram. We
use a reference (target) histogram to
provide a form for the specified histogram.
Histogram equalization is a special case of
histogram specification with a uniform reference
histogram.
* Gonzalez & Woods: Digital image processing. 3rd edn. (2007)
35. Page 35
Image Histogram
Def: Intensity histogram shows how many times a
particular intensity level appears in the image.
Mathematical description:
1. Non-normalized image histogram
2. Normalized image histogram – int. range levels
– row dimension
– column dimension
– the -th intensity value of the image
– number of pixel in the image with intensity
kk nrh )(
MNrrp kk
kr k
kn
kr
1,...,1,0 Lk
M
N
L
36. Page 36
Operations (Gonzalez & Woods, 2007):
1. Compute PDF (or a histogram) of an input image and
find the histogram equalization transformation. Round the
resulting values to the integer range .
Histogram Matching Algorithm
)( jr rp
k
s 1,0 L
k
0j
)()1()( jrkk rpLrTs
37. Page 37
2. Compute all values of the transformation function for
, where are the values of the
reference (target) histogram. Round the values of to
integers in the range . Store the values of in a
table.
Histogram Matching Algorithm
G
1,...,2,1,0 Lq )( iz
zp
1,0 L
q
i
izq zpLzG
0
)()1()(
G
G
38. Page 38
3. For each value of , use the stored values
of from step 2 above to find the corresponding values of
so that is closest to and store this mapping
from to :
,
.
Histogram Matching Algorithm
k
s 1...,,1,0 Lk
G
q
z )( qzG k
s
s z
kq
szG )(
)(1
kq
sGz
39. Page 39
4. Form the histogram-specified image in the following
manner:
4.1 First form histogram-equalized image of the input
image.
4.2 Then map every equalized pixel value of this image
to the corresponding value in the histogram-
specified image using the mapping from step 3 above.
Equalization step 4.1 above might be skipped by
combining two transformation functions and :
. .
Histogram Matching Algorithm
q
z
k
s
T 1
G
)())(( 11
kkq sGrTGz
41. Page 41
Image before Histogram Matching
*
Input image of the Mars moon and its non-
normalized histogram
* Gonzalez & Woods: Digital image processing. 3rd edn. (2007)
42. Page 42
Image after Histogram Matching
b *
* Gonzalez & Woods: Digital image processing. 3rd edn. (2007)
43. Page 43
Properties
Properties of the histogram matching:
a point operation
a “trial-and-error” process
considers the shape of the input and target
histograms
makes the actual histogram of the transformed
image to approximate the target histogram
44. Page 44
Properties
Properties of both histogram equalization and
matching:
point operations – no spatial information included
non-linear operations
the input, target and matched histograms have
equal intensity ranges (or length)
do not introduce new intensity values
45. Page 45
Histogram Matching
How it relates to our segmentation method:
a well-known algorithm for image enhancement
works well with images of slightly different intensity
ranges
adapts well to images with different types of
histograms
47. Page 47
Gray Level Co-occurrence Matrices
*
Gray Level Co-occurrence matrix (GLCM) of
horizontal neighboring pixel pairs in a matrix by
with an intensity scale .
* https://www.mathworks.com/help/images/ref/referenceetoh38.gif
4
8,1
5
48. Page 48
Gray Level Co-occurrence Matrix
GLCM is used in computer vision for:
texture analysis
(image classification)
feature extraction
image retrieval systems
object recognition
GLCM provides a useful estimation of similarity
between neighboring elements in a matrix/image.
49. Page 49
Our 2D Histogram
We use a 2D histogram as a sum of eight GLCMs,
corresponding to eight directions in a by window.
We apply it into two ways:
non-preprocessed, normalized (MR image
sequence division)
preprocessed, non-normalized (2D histogram
matching)
3 3
50. Page 50
Benchmark 2D Histogram
Another popular benchmark 2D histogram (Zhang
& Hu, 2008) has a bin with the following coordinates
and :
the first coordinate is an intensity value of a
central pixel, and
the second coordinate is an averaged intensity
value of eight pixels around the central one in the
local window by
i
j
i
j
3 3
51. Page 51
2D Histograms
our 2D histogram (preprocessed,
non-normalized) (set to zero the
first row and column bins)
“benchmark”
2D histogram
(Zhang & Hu, 2008)
52. Page 52
Properties
Our 2D histogram has:
comparatively good separability between the bins
(pixel pairs) of separate tissues and edges
good compactness and reflectional symmetry
along the diagonal
the bins of the tissue-background edges stay on
the first row and column in the 2D histogram
the pixel pairs of the background stay on bin
predominantly
)1,1(
53. Page 53
Properties
The second benchmark 2D histogram gained
popularity in the past because of:
its compactness and fewer outliers
the good discrimination of objects in the test
images from the s – few low-resolution images
its intuitive and fast computation
its less memory consumption
1980
54. Page 54
Properties
How it facilitates the matching and image
segmentation algorithms in our method:
compact distribution (no gaps in the areas of tissue
pixel pairs)
good discrimination between the pixel pairs of
tissues, edges and background
correspondence between the image pixel pairs and
the bins of the 2D histogram
56. Page 56
2D Histogram Matching
The position of the algorithm for 2D histogram
matching in our method:
compute a train matrix for each tissue of the test
MR image (partial classification) (using the 2D
histogram matching!)
kNN classify the test 2D histogram using the train
matrices (full classification)
segment the corresponding test MR image with the
classified test 2D histogram
57. Page 57
2D Histogram Matching
Important points of our implementation:
conversion of 2D histograms into vectors using a
zig-zag ordering
close but not equal length of the corresponding
test and model (target) vectors/histograms
truncation of the cumulative vectors within a
percentile interval
then truncation of the corresponding elements from
the normalized vectors
58. Page 58
2D Histogram Matching Algorithm
Operations of our implementation:
1. Compute test and model 2D histograms as a sum
of eight GLCMs within a 3x3 local window (eight
directions).
2. Convert 2D histograms into vectors using zig-zag
ordering.
3. Truncate the cumulative vectors within a percentile
interval. Truncate the corresponding elements
from the normalized vectors before matching.
59. Page 59
2D Histogram Matching Algorithm
4. Perform vector specification:
4.1 Calculate LUT between two cumul. vectors.
4.2 Calculate new values of the resulting vector
using the LUT between two normalized vectors.
5. Calculate the difference vector between the
resulting and initial normalized vectors.
6. Perform k-means clustering on the difference
vector, keep elements of the greatest cluster and
set others to zero.
7. Go back to the 2D histogram space and thus we
have a train matrix.
60. Page 60
Properties
Our 2D histogram matching algorithm:
uses position of pixel pairs distribution of different
tissues in the 2D histograms
considers correlation between the neighboring bins
in the 2D histogram by zig-zag ordering
localizes segments of bins of different tissues in
the 2D histogram as a final result
62. Page 62
Literature Review
https://i.ytimg.com/vi/FFcrQweRAp0/hqdefault.jpg
* https://i.ytimg.com/vi/FFcrQweRAp0/hqdefault.jpg
*
63. Page 63
Literature Review
Segmentation approaches of interest based on:
histograms, 2D histograms and related thresholds
models, built with different amount of information
different image spatial information
64. Page 64
Literature Review
Type of histograms used for image segmentation
1D histogram and related algorithms:
thresholding (Wu et al., 2007), find peaks (Morin et al., 2012)
and relative positions (Solanas et al.,2000), fit probability
distributions to а histogram form (Lee et al., 2009)
tissue intensity distributions as blind source separation
problem (BSS) (Zagorodnov & Ciptadi , 2011)
extraction of a feature vector from the histogram and
training a SOM classifier (Ortiz et al., 2013)
65. Page 65
Literature Review
2D histogram:
types of 2D histograms - 2D histogram (Kirby &
Rosenfeld, 1979), 2D-D histogram (Yimit et al., 2013), GLSC
histogram (Xiao et al., 2008), GLGM histogram (Xiao et al.,
2014)
kinds of thresholds – Otsu threshold (Zhang & Hu, 2008),
curvilinear threshold (Zhang & Hu,2009), entropic threshold
(Xiao, Cao, & Yuan, 2014), 2-phase threshold (Chen et al,
2010)
66. Page 66
Literature Review
Advantages of the suggested 2D histograms:
distinctive bins of tissues and edges
more information on edges in recent types of
2D histograms
lower memory consumption and higher speed of
computation
more accurate thresholds, easier and faster to
calculate
67. Page 67
Literature Review
Methods with different amount of training data:
unsupervised methods:
parametric models - GMM (Dong & Peng, 2014), RMM
(Roy et al., 2012)
non-parametric models - mean-shift (Mahmood, 2012),
kNN classification (Vrooman et al., 2013)
fuzzy models – FLGMM (Ji et al., 2012), FCM algorithm
(Ji et al., 2014)
68. Page 68
Literature Review
semi-supervised methods
ensemble framework (Azmi et al., 2013), Bayesian
transductive learning (Lee et al., 2013), spectral clustering
(Zhang et al.,2010)
supervised methods
atlas-based (Cabezas et al., 2011)
69. Page 69
Literature Review
Discussion:
the supervised methods are more accurate but
require more data that is labeled – a hurdle in the
current BIG Data era
the unsupervised methods are less accurate but
do not require labeled data – a great advantage
the semi-supervised methods are in-between …
–
70. Page 70
Literature Review
Methods which include image spatial information:
Markov random fields (MRF)
HMRF (Zhang et al., 2001), MWMAP (Monaco & Madabhushi,
2012), HMRF-MCMC ( Zhang et al., 2014)
variational methods
shape prior constraints (Liu et al., 2011), intensity
distributions (Chen & Radke, 2009)
level set theory methods
deformable models (Lee et al., 2012), INH artifact (Liu et al.,
2011)
71. Page 71
Literature Review
The methods use the following image properties:
piecewise smoothness between neighboring
pixels within tissues
piecewise discontinuity of neighboring pixels
between tissues
presence of noise and artefacts, which should be
included in the segmentation model
72. Page 72
Literature Review
How the above approaches are related to our
method:
Image spatial information, as it is included in the
test and model 2D histograms
Amount of preliminary information, as it selects
some MR images to build model 2D histograms
Transductive learning, as it performs 2D
histogram matching to find out the train matrices
Histogram thresholding, as the 2D histogram is
classified using the train matrices
73. Page 73
Summary
Points to remember:
MRI data – what are their characteristics,
artefacts, etc.
Transductive learning framework – how we
compute and apply our segmentation model
2D histogram – how we construct it
2D histogram matching – how we perform it
74. Page 74
Next - Part 2
1. Main idea and contributions
2. Introduction
3. Method description
4. Experimental results
5. Conclusions and future work
75. Page 75
Next - Part 2
3. Method description
3.1 Preprocess an MR image sequence
3.2 Divide into MR image subsequences
3.3 Compute test and model 2D histograms
3.4 Match a 2D histogram
3.5 Classify a 2D histogram
3.6 Segment using a back projection
76. Page 76
Next - Part 2
How will we merge 2D histograms and a 2D
histogram matching algorithm in a transductive
learning framework for a MR image segmentation?