MRI IMAGES THRESHOLDING FOR ALZHEIMER DETECTION

Sundarapandian et al. (Eds) : CSE, CICS, DBDM, AIFL, SCOM - 2013
pp. 95–104, 2013. © CS & IT-CSCP 2013 DOI : 10.5121/csit.2013.3310
MRI IMAGES THRESHOLDING FOR
ALZHEIMER DETECTION
Ali El-Zaart and Ali A.Ghosn
Department of Mathematics and Computer Science,
Beirut Arab University, Beirut, Lebanon
elzaart@bau.edu.lb , sawaghosn@hotmail.com
ABSTRACT
More than 55 illnesses are associated with the development of dementia and Alzheimer's disease
(AD) is the most prevalent form. Vascular dementia (VD) is the second most common form of
dementia. Current diagnosis of Alzheimer disease (Alzheimer's disease) is made by clinical,
neuropsychological, and neuroimaging assessments. Magnetic resonance imaging (MRI) can be
considered the preferred neuroimaging examination for Alzheimer disease because it allows for
accurate measurement of brain structures, especially the size of the hippocampus and related
regions. Image processing techniques has been used for processing the (MRI) image. Image
thresholding is an important concept, both in the area of objects segmentation and recognition.
It has been widely used due to the simplicity of implementation and speed of time execution.
Many thresholding techniques have been proposed in the literature. The aim of this paper is to
provide formula and their implementation to threshold images using Between-Class Variance
with a Mixture of Gamma Distributions. The algorithms will be described by given their steps,
and applications. Experimental results are presented to show good results on segmentation of
(MRI) image.
KEYWORDS
MRI Images, Image Thresholding, Between class variance, Gamma distribution.
1. INTRODUCTION
Alzheimer's disease (AD) is a progressive, degenerative disorder that attacks the brain's nerve
cells, or neurons, resulting in loss of memory, thinking and language skills, and behavioral
changes. As life span is increasing, an early detection of AD emerges as possible approach to
delay its consequences on patients and to increase the chance of getting potential benefits from
approved medications and maintain certain good level of life. In 2006, the worldwide prevalence
of AD was 26.6 million cases, and it was predicted that it will grow to 106.8 million in 2050.
Among these, there were 12.6 million cases in Asia (48% of worldwide prevalence), and it was
predicted that it will grow to 62.85 (59% of worldwide prevalence) in 2050.3
Routine structural neuroimaging evaluation is based on nonspecific features such as atrophy,
which is a late feature in the progression of the disease. Therefore, developing new approaches

96 Computer Science & Information Technology (CS & IT)
for early and specific recognition of Alzheimer disease at the prodromal stages is of crucial
importance.
Segmentation is an important issue in MRI image processing for diagnosis of (AD). One of the
simplest methods for image segmentation is thresholding [6]. It is based on selection of a value
that separates between two classes of gray-level values; this selected value is called threshold
value; which classify the gray-level value into two ranges to separate the object from the
background. The difficulty placed in determining the right threshold value or the optimal one.
Threshold techniques can be classified mainly into six groups according to the information they
are exploiting [8, 9, 10, 11, 12, 13, 14]. They are:
1) Histogram-based Methods. 2) Clustering-based Methods. 3) Object attribute-based Methods.
4) Spatial Methods. 5) Local Methods 6) Entropy-based Methods.
However, this work's goal is improving Otsu's method [6,7], so it can estimate the optimal
threshold values of images based on Gamma distribution, by using iteratively algorithm. Otsu’s
method is weak when it comes to dealing with low contrast images or where the object is small as
in [3]. In general, Otsu’s method produces a threshold value that maximizes between-class
variances. This work focuses on improving Otsu’s thresholding method to generate threshold
value automatically for a grayscale images based on Gamma distribution. Symmetric and
asymmetric histogram distribution of the intensity values can be represented by using Gamma
distribution [5]. Gamma distribution is extensively more than Gaussian distribution, which only
represents symmetric histogram distribution of the intensity values.
For each technique let ),( jif is the original image, T is the threshold value, and ),( jig is the
output image:


 ≥
=
otherwise,0
;),(for,1
),(
Tjif
jig
This paper will explore the related work in the field of image thresholding segmentation while
focusing on Otsu’s method. Section 2, demonstrates Gamma distribution. Our proposal method
explained in section IV. And in section V, experimental results show the efficiency of our method
in both bimodal and multilevel thresholding. Finally, in section VI will give the conclusion and
future works.
2. OTSU METHOD
Otsu’s method [1] is one of the most popular methods for its simplicity and efficiency [4]. It is
based in thresholding technique. It depends on selecting the optimal threshold value that
maximizes the between-class variance of resulting object and background classes. The search for
the optimal threshold done sequentially until finding a value that makes variance between two
classes or more maximum. In this section, will demonstrate Otsu’s method and its development
which made it fast in computational. If the image in the two dimensions is represented by the
function f (x, y) and the values of it in gray-level have the range between [0...L], where L=255.
And let N represent the whole number of pixels in the image. The number of pixels with gray-
level i ish(i), i= [0,1,..., L] that represent the histogram of the image. For the simplest, normalize

Computer Science & Information Technology (CS & IT) 97
of the histogram was computed which represents the probability ofoccurrence of gray-level i as
the follows:pሺiሻ =
୦ሺ୧ሻ
୒
, pሺ݅ሻ ≥ 0
The total average or mean value of the image computed as:
µ்
= ෍ ipሺ݅ሻ
୐ିଵ
௜ୀ଴
෍ pሺ݅ሻ
୐ିଵ
௜ୀ଴
= 1
For the single thresholding, image pixels will be divided into two classesC1 {0,1,..., t}, and
C2{t,,..., L-1}, where tis the threshold value. Usually, the resulted image from this method
corresponds to separate the object (C2thatrepresents the class of bright pixels) from the
background (C1that represents the classof dark pixels). The probability of these two classes is:
߱ଵሺ‫ݐ‬ሻ = ∑ ‫݌‬ሺ݅ሻ௧ିଵ
௜ୀ଴ and߱ଶሺ‫ݐ‬ሻ = ∑ ‫݌‬ሺ݅ሻ௅
௜ୀ௧
And the mean for them are:
µଵ
ሺ‫ݐ‬ሻ = ෍ ݅ℎሺ݅ሻ
௧
௜ୀ଴
/ωଵሺ‫ݐ‬ ሻ
µଶ
ሺ‫ݐ‬ሻ = ෍ ݅ℎሺ݅ሻ/ωଶሺ‫ݐ‬ ሻ
௅ିଵ
௜ୀ௧
Then by using discriminate analysis [1], Otsu proves that optimal threshold (t*)can be obtained
by maximizing the between-class variance as:
t∗
= Arg max଴ஸ୲ஸଶହହ{σ୆
ଶ ሺtሻ} (1)
Where the between-class variance (σ୆
ଶ ሺtሻ ) is defined as:
ηሺtሻ = σ୆
ଶ ሺtሻ
σ୆
ଶ ሺtሻ = ωଵሺtሻሺμଵሺtሻ − μ୘ሻଶ
+ ωଶሺtሻሺμଶሺtሻ − μ୘ሻଶ
(2)
This method also has been extended to multilevel thresholding of image by Otsu [1]. If we have
M classes where:
C1 = {0, .., t1}, C2 = {t1,.., t2} and CM = {tM-1 + 1,.., 255 }
Then the threshold values will be:
{‫ݐ‬ଵ
∗
, ‫ݐ‬ଶ
∗
, … … ‫ݐ‬୑ିଵ
∗ } = Arg max
଴ஸ୲ஸଶହହ
{ ෍ ߱௞
ெ
௞ୀଵ
ሺߤ௞ − ߤ்ሻଶ
}
Many papers have been published which study the improvement of this method [7, 21]. In [2], the
author improves Otsu’s method to make it more efficient in computational side. He used the
derivation computation to find the optimal threshold value in iteratively manner rather than

searching in sequential that decrease the loop, and then the time of CPU needed for the
computational process. He concluded to the following formula for two classes:
‫ݐ‬ =
ߤଵ + ߤଶ
2
Where µ1is the mean of the first class and µ2 is the mean for second class. This formula has
convergence property [56] which means that in each iteration the threshold value converges to be
optimal.
In [7], the authors also developed algorithm to compute multilevel threshold values faster, using
recursive algorithm with look-up table that shortens the huge number of mathematical operations
needed in this situation.
However, this method is only applied when the histogram is symmetric but for asymmetric
histogram it is better to use Gamma distribution as described in the next section.
3. GAMMA DISTRIBUTION
Histogram represents statistical information for image pixels. It describes pixels intensity
distribution in an image by graphing the number of pixels intensity at each gray level or color
intensity level. Essentially, we can say that there are two types of distributions of gray level
(mode): symmetric and non-symmetric. For the symmetric mode Gaussian distribution work well
to estimate the threshold value. On the other hand, Gamma distribution is more general to
represent the both symmetric and non-symmetric modes. The Gamma function defines as [5]:
݂ሺ‫,ݔ‬ ߤ, ܰሻ =
ଶ௤
ఓ
ேಿ
௰ሺேሻ
[
௤௫
ఓ
]ଶேିଵ
݁
ିேሺ
೜ೣ
ഋ
ሻమ
(3)
Where q = ΓሺN + 0.5ሻ/√NΓሺNሻ,x is the intensity of the pixel, µ is the mean value of the
distribution and N is the shape of distribution. In our method, Gamma distribution used to
estimate the mean values of the image modes and then find the optimal threshold value. Figure 1,
shows the Gamma distribution for one mode with different shape parameter N and same value of
mean µ.
Figure 1: Gamma Distribution with the same mean and different shape parameter N.
4. VALLEY-EMPHASIS METHOD USING A MIXTURE OF GAMMA
DISTRIBUTIONS FOR SMALL OBJECT
This Section will demonstrate our proposal method. It solves the problem of image segmentation
in case if it contains small object. As mentioned before in section 3, this method also based
onOtsu’s method using Gamma distribution. We improved our method and gave it the ability to
deal with small objects. As well as it can deal with asymmetric mode in the histogram to get more

satisfying segmentationresults. We give comprehensive explain on the bimodal and multimodal
thresholding algorithms.
4.1. Bimodal Thresholding
Hui-Fuang suggested adding a weight with Otsu’s method for detecting small object [3]. The
author observes when select the threshold value to be on the valley of the two peaks (from the
histogram), the probability of occurrence at the threshold value (pt) has to be small. Based on this
observation he proposed to add the weight to Otsu’s method which will improve it for selecting
threshold values. However, this method called “valley-emphasis”.Objective to select a threshold
value that has small probability of occurrence (valley in the gray-level histogram), and it also
maximizes the between class variance, as in the Otsu method [3]. By applying the weight (1-pt) to
Otsu’s method is the key of the valley-emphasis method. The smaller the (p(t)) value, the larger
the weight (1-pt) will be. This weight ensures that the result threshold value will always be
located in at the valley or bottom rim of the gray-level distribution [3].As Otsu’s method, the
valley-emphasis method also attempts to maximize the between-class variance of the histogram.
Based on Hui-Fuang’s work we proposed our method for bimodal thresholding as the following:
ηሺtሻ = σ୆
ଶ ሺtሻ = [ωଵሺtሻሺμଵሺtሻ − μ୘ሻଶ
+ ωଶሺtሻሺμଶሺtሻ − μ୘ሻଶ]ሺ1 − pሺtሻሻ
(4)
Where the probability of occurrence of gray-level t is defined as:
‫݌‬ሺ‫ݐ‬ሻ =
୦ሺ୲ሻ
୬
; P′
ሺ‫ݐ‬ሻ =
୦ሺ୲ሻି୦ሺ୲ିଵሻ
୬
n is the total number of pixels in a given image. P’(t)is the first derivative of the probability.
We consider h(i), i = 0 … 255 be the bimodal histogram of the original image. We assume that
the histogram of an image can be seen as a combination of Gamma distributions. The mean
values of each mode can be estimated using Gamma distribution [5]. We used Gamma
distribution because it has the ability to represent both symmetric and non-symmetric mode rather
than the limited Gaussian distribution that describes only the symmetric mode better. However,
the mean values will be as the following:
ߤଵሺ‫ݐ‬ሻ = ට
∑ ௛ሺ௜ሻ೟షభ
೔సబ ௜మ௤మ
∑ ௛ሺ௜ሻ೟షభ
೔సబ
=
ට∑ ௛ሺ௜ሻ೟షభ
೔సబ ௜మ௤మ
ඥఠభሺ௧ሻ
=
ఓభೌሺ௧ሻ
ఓబೌሺ௧ሻ
(5)
ߤଶሺ‫ݐ‬ሻ = ට
∑ ௛ሺ௜ሻಽ
೔స೟ ௜మ௤మ
∑ ௛ሺ௜ሻಽ
೔స೟
=
ට∑ ௛ሺ௜ሻಽ
೔స೟ ௜మ௤మ
ඥఠమሺ௧ሻ
=
ఓభ್ሺ௧ሻ
ఓబ್ሺ௧ሻ
(6)
Where µT the total mean of image, µ1 the mean value of the first class and µ2 the mean of the
second class.h(i), I = 0 … 255 is the histogram of image. Here, we aim to divide the histogram in
two classes C1{0,1,..., t}, and C2{t+1,..., 255}. Where t is the threshold value. Consequently we
can define the thresholded image as:
gሺx, yሻ = ൜
0 fሺx, yሻ ≤ t
255 fሺx, yሻ > ‫ݐ‬
(7)

Where fሺx, yሻ, is the original image, and gሺx, yሻ is the segmented image result.
Therefore, we calculate the first derivative of Eq. (4) then we set it to zero [2] to get the optimal
threshold value as:
[߱ଵሺ‫ݐ‬ሻሺߤଵሺ‫ݐ‬ሻ − ߤ்ሻଶ
+ ߱ଶሺ‫ݐ‬ሻሺߤଶሺ‫ݐ‬ሻ − ߤ்ሻଶ]′
ሺ1 − ‫݌‬ሺ‫ݐ‬ሻሻ + [߱ଵሺ‫ݐ‬ሻሺߤଵሺ‫ݐ‬ሻ − ߤ்ሻଶ
+
߱ଶሺ‫ݐ‬ሻሺߤଶሺ‫ݐ‬ሻ − ߤ்ሻଶ]ሺ‫݌‬ሺ‫ݐ‬ሻ − 1ሻ′
= 0The first derivative of Eq. (2) is:
ൣ࣓૚ሺ࢚ሻሺࣆ૚ሺ࢚ሻ − ࣆࢀሻ૛
+ ࣓૛ሺ࢚ሻሺࣆ૛ሺ࢚ሻ − ࣆࢀሻ૛
൧
′
= [߱ଵ
′ ሺ‫ݐ‬ሻሺߤଵሺ‫ݐ‬ሻ − ߤ்ሻଶ
+ 2 ߱ଵሺ‫ݐ‬ሻሺߤଵሺ‫ݐ‬ሻ −
ߤ்ሻߤଵ
′
+ ߱ଶ
′ ሺ‫ݐ‬ሻሺߤଶሺ‫ݐ‬ሻ − ߤ்ሻଶ
+ 2 ߱ଶሺ‫ݐ‬ሻሺߤଶሺ‫ݐ‬ሻ − ߤ்ሻߤଶ
′ ሺ‫ݐ‬ሻ] (8)
߱ଵ
ᇱ ሺ‫ݐ‬ሻ = ℎሺ‫ݐ‬ሻ ; ߱ଶ
ᇱ ሺ‫ݐ‬ሻ = −ℎሺ‫ݐ‬ሻ (9)
ߤଵ
ᇱ ሺ‫ݐ‬ሻ =
௛ሺ௧ሻ
ଶ ఓబೌሺ௧ሻ
ቂ
௧మ௤మ
ఓభೌሺ௧ሻ
−
ఓభሺ௧ሻ
ఓబೌሺ௧ሻ
ቃ (10)
ߤଶ
ି௛ሺ௧ሻ
ଶ ఓబ್ሺ௧ሻ
ቂ
௧మ௤మ
ఓభ್ሺ௧ሻ
−
ఓమሺ௧ሻ
ఓబ್ሺ௧ሻ
ቃ (11)
ߤଵ௔
௛ሺ௧ሻ௧మ௤మ
ଶ ఓభೌሺ௧ሻ
; ߤ଴௔
௛ሺ௧ሻ
ଶ ఓబೌሺ௧ሻ
ߤଵ௕
ᇱ ሺ‫ݐ‬ሻ = −
௛ሺ௧ሻ௧మ௤మ
ଶ ఓభ್ሺ௧ሻ
; ߤ଴௕
ᇱ ሺ‫ݐ‬ሻ = −
௛ሺ௧ሻ
ଶ ఓబ್ሺ௧ሻ
Equation (8) will lead to
A(t) + B(t) t2
= 0 (12)
Where
Aሺtሻ = hሺtሻሺn − hሺtሻሻμ୘൫μଶሺtሻ − μଵሺtሻ൯ − [hሺtሻ − hሺt − 1ሻ][ωଵሺtሻሺμଵሺtሻ − μ୘ሻଶ
+
ωଶሺtሻሺμଶሺtሻ − μ୘ሻଶ
](13)
‫ܤ‬ሺ‫ݐ‬ሻ = ℎሺ‫ݐ‬ሻ‫ݍ‬ଶሺ݊ − ℎሺ‫ݐ‬ሻሻ[
૚
ࣆ૛ሺ࢚ሻ
−
૚
ࣆ૚ሺ࢚ሻ
] (14)
∵ ‫ܣ‬ሺ‫ݐ‬ሻ + ‫ܤ‬ሺ‫ݐ‬ሻ‫ݐ‬ଶ
= 0
∴ ‫ݐ‬ = + ට
ି஺ሺ௧ሻ
஻ሺ௧ሻ
; (15)
B(t) ≠ 0 and
ି୅ሺ୲ሻ
୆ሺ୲ሻ
≥ 0
From the Eq. (15), we will get the optimal threshold value (t).
This formula allows us to use an iterative thresholding algorithm that makes the new threshold
value converge to the optimal threshold value and decrease the processing time of calculation.
The bimodal thresholding algorithm of thismethod is described in as:

4.2. Multimodal Thresholding
In multilevel thresholding, the resulted image will be segmented to M mode according to a set of
threshold values T , where T= {t1, t2, t3, ….., tM-1}, where:
t0 = 0 < t1< t2< t3< ……<tM-1<tM =255. The means of each mode (classes) or the kth mode
defined as in Eq. (16).
µ୩
ሺtሻ = ඨ
∑ ୦ሺ୧ሻ
౪ౡ
౟స౪ౡషభ
୧మ୯మ
∑ ୦ሺ୧ሻ
౪ౡ
౟స౪ౡషభ
ሺ16ሻ
k= 1, 2 ……, M
The optimal threshold value T* = {tଵ
∗
, tଶ
∗
, tଷ
∗
, … . , t୑ିଵ
∗
,} for each class will be estimated by using
the bimodalthresholding mention in section 4.A. The multilevel thresholding algorithm is
described as:

5. EXPERIMENTAL RESULTS
This section provides experimental results for applying our valley-emphasis method on MRI
images. Figure 2 shows an original MRI images and the segmentation result.

Figure 2. Original MRI images and segmented Results.
6. CONCLUSIONS
We have proposed a new method for image thresholding (bimodal and multimodal thresholding)
based on between-class variance and using Gamma distribution. Our method solves the problem
of non-symmetric histogram of images by using Gamma rather than Gaussian distribution. We
have also applied it iteratively to make the threshold value converges to be optimal value, and to
reduce the computations. In addition, we provided experimental results that show our method
efficiency in comparison with the original Gaussian Otsu method. In addition we extend our
proposed method to solve the problem for detecting the small object in the images. As well as
shown in the experimental results we have satisfied segmentation results. Finally we applied our
algorithm in many other images as medical images.
REFERENCES
[1] N.Otsu, “A threshold selection method from gray-level histogram,” IEEE Transactions on Systems,
MAN, and Cybernetics, Vol. SMC-9, No. 1, 1979, pp. 62-66.
[2] K.Chin Lin, “Fast thresholding computation by searching for zero derivatives of image Between-
Class Variance,” the 27th Annual Conference of the IEEE Industrial Electronics Society, 2001, pp.
393-397.
[3] H.Fuang Ng, "Automatic Thresholding for Defect Detection," Pattern Recognition Letters, Vol. 27,
2006, pp.1644-1649.
[4] Z.Hou, Q. Hu and W.L. Nowinski, “On minimum variance thresholding,” Pattern Recognition
Letters, Vol. 27, 2006, pp. 1732-1743.
[5] A.El Zaart, D. Ziou, S. Wang and Q. Jiang, " Segmentation of SAR images,” Pattern Recognition
Journal, Vol. 35, No. 3, 2002, pp. 713-724.
[6] J.C. Russ, “The image processing handbook,” CRC Press, Boca Raton, NW, USA, 5th edition, 2007.
[7] P.Liao, T. Chen, and P. Chung, “A Fast Algorithm for Multilevel Thresholding,” Journal of
Information Science and Engineering, Vol. 17, 2001, pp. 713-727.

[8] T.Logeswari ,M. Karnan (2010). An improved implementation of brain tumor detection using
segmentation based on soft computing. J. Cancer Res. Exp. Oncol., 2(1): 006-014.
[9] K.Mirzaei, H. Motameni, R. Enayatifar (2011). New method for edge detection and de noising via
fuzzy cellular automata. Int. J. Phy. Sci. 6(13): 3175-3180.
[10] P.Sathya, R. Kayalvizhi(2010). PSO-Based TsallisThresholdingSelection Procedure for Image
Segmentation. Int. J. Comput. Appl.5(4): 39-46.
[11] O.Sojodishijani, V .Rostam, A. Ramli (2010). A video-rate color image segmentation using adaptive
and statistical membership function. Sci. Res. Essays, 5(24): 3914-3925.
[12] L.Yu (2008). Vehicle extractions using histograms and genetic algorithm based fuzzy image
segmentation from high resolution UAV aerial imagery. The International Archives of the
Photogrammetric, Remote Sens. Spat.
[13] E.Arias-Castro, DL. Donoho (2009). Does median filtering truly preserve edges better than linear
filtering? Annals of Statistics. 37(3): 1172–2009.
[14] Y.Fernandez, A. Sarinana, R. Swenson (2009). development of a prototype for classification of
potato mini-tubers based on artificial vision. 6th International Conference on Electrical Engineering,
Computing Science and Automatic Control.
AUTHOR
Ali El-Zaart was a senior software developer at Department of Research and Development, Semiconductor
Insight, Ottawa, Canada during 2000-2001. From 2001 to 2004, he was an assistant professor at the
Department of Biomedical Technology, College of Applied Medical Sciences, King Saud University. From
2004-2010 he was an assistant professor at the Department of Computer Science, College
of computer and information Sciences, King Saud University. In 2010, he promoted to
associate professor at the same department. Currently, his is an associate professor at the
department of Mathematics and Computer Science, Faculty of Sciences; Beirut Arab
University. He has published numerous articles and proceedings in the areas of image
processing, remote sensing, and computer vision. He received a B.Sc. in computer science
from the Lebanese University; Beirut, Lebanon in 1990, M.Sc. degree in computer science
from the University of Sherbrooke, Sherbrooke, Canada in 1996, and Ph.D. degree in
computer science from the University of Sherbrooke, Sherbrooke, Canada in 2001. His research interests
include image processing, pattern recognition, remote sensing, and computer vision.

MRI IMAGES THRESHOLDING FOR ALZHEIMER DETECTION

More Related Content

What's hot (20)

Similar to MRI IMAGES THRESHOLDING FOR ALZHEIMER DETECTION (20)

More from cscpconf (20)

Recently uploaded (20)

MRI IMAGES THRESHOLDING FOR ALZHEIMER DETECTION