PREPRINT 2012:18

On Efficiency of Combined
Daubechies Wavelets and Statistical
Parameters Applied in Mammography

M. ASADZADEH
E. HASHEMI
A. KOZAKEVICIUS

Department of Mathematical Sciences

Division of Mathematics
CHALMERS UNIVERSITY OF TECHNOLOGY
UNIVERSITY OF GOTHENBURG
Gothenburg Sweden 2012
 Preprint 2012:18

On Efficiency of Combined Daubechies Wavelets

and Statistical Parameters Applied
in Mammography

M. Asadzadeh, E. Hashemi and A. Kozakevicius

Department of Mathematical Sciences

Division of Mathematics
Chalmers University of Technology and University of Gothenburg
SE-412 96 Gothenburg, Sweden
Gothenburg, September 2012
 Preprint 2012:18
ISSN 1652-9715

Matematiska vetenskaper
Göteborg 2012
 ON EFFICIENCY OF COMBINED DAUBECHIES WAVELETS AND
STATISTICAL PARAMETERS APPLIED IN MAMMOGRAPHY

M. ASADZADEH1,† , E. HASHEMI 1,2

AND A. KOZAKEVICIUS 1,3,‡,∗

Abstract. This note is a study of combined Daubechies (Db2) wavelet transform and the
statistical parameters skewness and kurtosis applied to detection of microcalcification in mam-
mography. We have succintly introduced the concept of vanishing moments and derived scaling
and wavelet functions using generating functions. The efficiency of the discrete algorithm is
heavily relied on the order of performing wavelet approximation and the statistical procedure.
The vital importance of both wavelet and statistical parameter approaches as well as the order-
ing issue in performing the analysis are justified through implementing numerical examples for
some clinical data.

Keywords: Daubechies wavelet, vanishing moments, mammography, microcalcification, skew-

ness, kurtosis.

AMS Subject Classification: 65T60, 68U10, 97K80, 97M60.

1. Introduction
Mammography is a common expression for specific type of imaging that uses a low-dose
x-ray (photon-beam) therapy to examine breast tissue and a mammography exam is called a
mammogram [13]. One of the main indicators of breast cancer searched in mammograms is a
set of clusters formed by microcalcifications, these are tiny calcium deposits in breast tissue,
that appear as small bright spots in the imaging [13].
In the last 15 years several mammography/image processing methods have been developed in
order to help radiologists with the task of detecting microcalcifications. Among them wavelet
based methods have been designed also in association with different statistical measurements
[11, 12]. According to [12], microcalcifications in mammograms correspond to high frequency
coefficients of the image spectrum, and a possible procedure to detect and extract these calcifi-
cations is simply to decompose the mammography image by wavelet transforms [2], suppress the
low frequency subband (scaling coefficients block), and reconstruct a new image considering only
the high frequency wavelet coefficients. The draw-back in this approach is that such procedure
leads to a high number of false positive results, since no distinction is done among the high
intensity wavelet coefficients.
To filter out false positive results and identify the regions of interest a useful tool is to employ
the statistical parameters, called skewness and kurtosis. This is in the sense that, in a region
containing microcalcifications, the symmetry of the Gaussian distribution of wavelet coefficients
is deteriorated and hence the tails of their distribution are heavier, and/or their heights are
‡
The corresponding author.
1
Chalmers University of Technology, Sweden, 2 Polytechnique de Torino, Italy, 3 Universidade Federal de Santa
Maria- UFSM, Brazil,
e-mail:mohammad@chalmers.se, eliza.hashemi@gmail.com, alicek@ufsm.br
†
The research of this author is supported by the Swedish Foundation of Strategic Research (SSF) in Gothenburg
Mathematical Modeling Center (GMMC) and Swedish Research Council (VR).
The author∗ would like to thank the National Counsel of Technological and Scientific Development- CNPq (Brazil)
under grant No 201457/2010-5 and the Mathematical Sciences Department of Chalmers University of Technology
and University of Gothenburg.
Manuscript received xx.
1
peaked or plat [11]. The statistical parameters measuring these deformation effects in Gaussian
distributions are the third and fourth order correlation parameters, called skewness and kurtosis,
respectively. Therefore their computation for the wavelet coefficients may be used to characterize
regions most similar to microcalcifications clusters, as the regions of interest. The analysis of
these parameters, in connection with wavelet transforms goes back to [12], where the first analysis
were done considering undecimated wavelet transforms. The ideas initially presented in [12] were
then revisited in [11].
The main differences between the proposed approach in here and the methodology presented
in [11] is that, now the discrete Daubechies wavelet transform with two vanishing moments,
Db2, is employed, leading to simpler analysis algorithms, with the advantage of minimizing
the boundary distortions caused by long sized filter wavelet functions, as for example wavelets
with a higher number of vanishing moments. With the Db2 filters it is possible to detect and
precisely localize continuous piecewise linear signals. A property that is carried out also in the
two dimensional context, i.e. to detect and localize subregions inside the image corresponding to
linear data variations. For such a signal or image Db2 would yield zero wavelet coefficients (once
the linear information is entirely represented by the scaling coefficients due to the two vanishing
moments) at the linearity subregions. Nevertheless at the site of singularities and some close
neighborhoods the wavelet coefficients would capture the higher variations, which due to the
type of singularity, can be spread out for many levels of the transformation.
Another difference of the proposed algorithm and the previous studies is that now there is
no need to consider any overlapping subregions. This change implies in a major simplification
of the microcalcification detection procedure, since no strategy for selecting and dealing with
the subregions analysis is necessary. Instead, in our algorithm, three subbands with the corre-
sponding wavelet coefficients are considered, where on each row and column of these subbands,
the statistical parameters skewness and kurtosis are computed. The vectors containing these
statistical quantities, computed by rows and columns, are then thresholded, keeping only the
significant values, those which are higher than a chosen threshold parameter (here 80% of the
maximum of the details of each band). The crossing of rows and columns associated to the
significant values determine candidate regions of microcalcifications clusters.
An outline of this work is as follows. In Section 2 a swift framework to the Daubechies
wavelet (Db2) construction is summarized, highlighting the importance of the vanishing moments
property. We have also introduced some aspects of the statistical parameters: skewness and
kurtosis. The analysis of the algorithm is presented in Section 3. Also a series of tests are
designed in order to validate the methodology and to stress issues related to the relevance of
the combination of the two chosen parameters with the wavelet transform. Numerical results,
presented in Section 4, are based on the analysis of mammography from the web page of the
American cancer society. The main advantage of considering data from this data base is that
annotations from specialists are provided with data, which is primordial for the validation of
the developed algorithms. Finally, Section 5 is devoted to comments and concluding remarks.

2. Wavelet framework and Statistical Parameters

2.1. The Daubechies wavelet DB2. In this section we wish to motivate the analytical aspects
considered in the choice of the Daubechies Wavelet with two vanishing moments (Db2) in our
study of mammography. A wast class of wavelet families suffers from some major drawbacks
which can lead to some inconveniences during the analysis. Haar wavelets have compact support
but are discontinuous. Linear spline wavelets are continuous, but the orthogonal scaling function
and associated wavelet have infinite support, they do, however, decay rapidly at infinity, [6].
The hierarchy of Daubechies wavelet family contains the Haar wavelet, which is the only
discontinuous one. The other orthonormal Daubechies wavelets are both compactly supported
and continuous. The degree of smoothness of Daubechies wavelets increase with raising their
 englishMICROCALCIFICATION DETECTION USING WAVELETS 3

hierarchy number, associated to the capacity the scaling functions have to span polynomials
from degree 0 until a certain degree n − 1. The number of vanishing moments of the wavelet
family is then said to be n, since the wavelet coefficients for those n smooth functions are all zero,
for all levels of decomposition (by the orthogonality between wavelets and scaling functions) [5].
To detect a biological disorder in human tissue, we are dealing with certain biochemical models
in a bounded three dimensional domain with particular singularity aspects. On the other hand
a raw model for human tissue/organ would involve differential equations and hence requires
to assume a certain degree of smoothness. The main computational objective of this study is
two-fold: i) To model a problem setting using optimally small number of degrees of freedom in
order to keep the cost down. ii) To single out the healthy tissue regions from the regions of
interest in order to help the oncologist in radiation treatment of malignant tumor. Therefore,
here Db2 fits well for the purpose of our study, specially because the considered data is discrete
and the way boundaries are treated matters now for the final analysis.
Below we give a brief mathematical background for the construction of Db2. In this setting,
to introduce Db2 wavelets we state the properties that guarantee the pointwise and L2 conver-
−iξ
gence of scaling function defined by a polynomial p(ξ) = P (e−iξ ), (with P (e−iξ ) = 1+e2 =
e−iξ/2 cos(ξ/2)). Obviously p(ξ) satisfies the properties

 p(0) = 1
|p(ξ)|2 + |p(ξ + π)|2 = 1 (1)
|p(ξ)| > 0, for |ξ| ≤ π/2.


We recall the scaling equation

X
φ(x) = pk φ(2x − k),
k
wher the coefficients pk are given by
Z ∞
pk = 2 φ(x)φ(2x − k) dx.
−∞

Now let φ0 (x) be the Haar scaling function and define

X
φn (x) = pk φn−1 (2x − k), for n ≥ 1, (2)
k∈Z

where, Z is the set of integers. Then the sequence {φn } converges pointwise and in L2 to a
function φ̃, satisfying the orthogonality condition
Z ∞
φ̃(x − n)φ̃(x − m) dx = δnm , (3)
−∞
P
and the scaling equation φ̃(x) = k pk φ̃(2x − k). Obviously φ̃ ≡ φ. The function
1 + e−iξ
P (e−iξ ) = = e−iξ/2 cos(ξ/2),
2
is called the generating polynomial and is an example of a polynomial satisfying (1). The
iterative procedure above yields the Daubechies scaling function φ, and the associated wavelet
is given by
X
ψ(x) = (−1)k p1−k φ(2x − k). (4)
k∈Z
A general form for the generating polynomial is given by
1X
P (z) = pk z k , (5)
2
k∈Z
4 english M. ASADZADEH1,† , E. HASHEMI 1,2 AND A. KOZAKEVICIUS 1,3,‡,∗

where z is a complex number with |z| = 1. A common truncation is of the form PN (z) =
1 P2N −1 k
2 k=0 pk z , where in the two simplest cases with N = 1 (Haar) and N = 2 (Daubechies), the
polynomials are
1  1 + √3 1 − √3 
P1 (z) = (1 + z), and P2 (z) = + z (1 + z)2 . (6)
2 8 8
In general we have the following result.
Proposition 2.1.1 ( Daubechies). For every N , there will be 2N nonzero, real scaling filter
coefficients p0 , p1 , . . . , p2N −1 , resulting in a scaling function and wavelet that are supported on
the intervalP0 ≤ t ≤ 2N − 1. They are chosen so that the corresponding degree 2N − 1 polynomial
2N −1
PN (z) = 21 k=0 pk z k has the factorization
PN (z) = (z + 1)N P̃N (z), (7)
where the degree of P̃N is N − 1 and P̃N (−1) 6= 0.
This guarantees that the associated wavelets will have precisely N vanishing moments.
1
Thus, both polynomials P1 (z) and P2 (z) in (6) have the factorization (7) with P̃1 (z) = 2 and
√ √
1+ 3 1− 3
P̃2 (z) = 8 + 8 z. Since PN are real, Pn (−z) = Pn (−z̄), and one can easily show that the
scaling function φN and wavelet ψN have Fourier transforms given by
1 iξ/2j
φ̂N (ξ) = √ Π∞j=1 PN (e ), (8)
2π
ψ̂N (ξ) = e−iξ/2 PN (eiξ/2 )φ̂N (ξ/2). (9)
Then, we have that

(k) 0 √ k = 0, . . . , N − 1,
ψ̂N (0) = (10)
−N !(−i/2)N P̃ N (−1)/ 2π 6= 0, k = N.
This yields the following result by Daubechies:
Proposition 2.1.2 (Daubechies). The wavelet function ψN possesses exactly N vanishing mo-
ments, i.e., we have
Z ∞ 
√0 k = 0, . . . , N − 1,
xk ψN (x) dx = −N N !/ 2π)P̃ (−1), k = N. (11)
−∞ −(2 N

Daubechies wavelets are classified according to the number of vanishing moments they have.
The smoothness of the scaling function and wavelet increases with their number of vanishing
moments. As we mentioned above, the case N = 1 is the same as the Haar case, where both
scaling function and wavelet are discontinuous. For N = 2, the Daubechies scaling function and
wavelet are continuous but certainly do not have smooth derivatives. In the N = 3 case, both
scaling and wavelets functions are continuously differentiable. For more details see, e.g. [3, 1].
Why is it useful to have vanishing moments? The short answer is that vanishing moments are
a key factor in many applications-compression, noise removal, singularity detection, for example.
According to expression (11) for the moments computation, in the case N = 2 (Db2), the first
two moments (k = 0, 1) vanish, and no information is represented in terms of the wavelet, just
with respect to the scaling function. For the numerical applications this implies that only half
of the discrete data, exactly the scaling coefficients, is necessary to represent the original signal,
since all wavelets coefficients will be zero, as indicated by (11).
√ √ √
For polynomial of degree k = 2 considering Db2, −(2−1 / 2π)P̃2 (−1) = −1/(2 2π) · 3/4 is
the corresponding computed moment. Thus for the third moment we have
Z ∞ r
2 1 3
x ψ2 (x) dx = − . (12)
−∞ 8 2π
 englishMICROCALCIFICATION DETECTION USING WAVELETS 5

We can use these moments to approximate and estimate the wavelet coefficients for smooth
signals and show that these coefficients will be small when the level j is high. If f is a smooth,
twice continuously differentiable signal, then for sufficiently large j, using Taylor expansion, its
j, k-wavelet coefficient is approximated as,
Z ∞ Z 3.2−j
j
dk = j/2 j
f (x)2 ψ2 (2 x − k) dx = f (x + 2−j k)2j/2 ψ2 (2j x) dx
−∞ 0
Z 3.2−j  (13)
−j 0 −j 1 2 00 −j  j/2 j
≈ f (2 k) + xf (2 k) + x f (2 k) 2 ψ2 (2 x) dx.
0 2
Now, since the first two moments vanish and the third is given in (12), we can approximate the
wavelet coefficients for any level j and any position k as
r
j 1 3 −5j/2 00 −j
dk ≈ − 2 f (2 k). (14)
16 2π
Note that the formula (14) is exact in any region where the signal (the function f ) is a polynomial
of degree ≤ 2 in x. As an application of this formula, we find a point where an otherwise smooth
function has a discontinuity in the derivative. This application is called singularity detection,
and this process can be used, among other things, to the purposes of interest in this work: the
detection of microcalcification regions in mammogram images.
As an example, consider a piecewise linear function defined in a compact interval, with a
slope change at a single interior point in the interval. Since f 00 ≡ 0 where f is linear, hence (14)
implies that the only nonzero wavelet coefficient will come from a small region near the corner
point where the slope changes.
Example 2.1.3. Consider the piecewise linear signal

0.5x + 1, 2≤x≤3
f (x) =
1.5(x − 1), 3<x≤4
After sampling the signal at 128 equally spaced points, we perform one level of decomposition
using the Db2 (N = 2 Daubechies) wavelet with 2N scaling filters. Thus our starting finest level
of representation is indexed by j = 7, and j = 6 is the level immediately coarser obtained by the
transform. The only two considerable non-zero wavelet coefficients are dj30 and dj31 corresponding
to a wavelet supported on the interval [3 − αhl , 3 + αhl ], where α = 2N is the size of the scaling
filters and hl = lh0 is the corresponding spacing according to the level of decomposition, here
l = 1 and the finest grid spacing for the original grid is h0 = 1/64. Consequently, the singularity
is supported in this interval.
2.2. Skewness and Kurtosis. A thorough wavelet study would yield a final imaginging that
does not miss any obnormalities. This would include both malingnant tumors as well as healthy
binals. To detect the true positives, the sort of false positives are heavily relied on the match
between the two statistical measuring parameters: skewness and kurtosis. More specifically, the
need to study skewness and kurtosis is due to the fact that the normal curve often fails to give
an adequate representation for a considered data.
These quantities, which are playing a crucial role in many applications with data interpreta-
tions, are not usually described mathematically. Below we give a short description of the concepts
of skewness and kurtosis and refer the reader to standard statistical texts for further properties
and some detailed descriptions. The need to study skewness and kurtosis is due to the fact that
the normal curve often fails to give an adequate representation for a considered data. While
skewness is a measure of the asymmetry of the probability distribution function, the kurtosis
can be employed as a measure of its peakedness. In Pearson’s curve system it has been suggested
that to model the data selected on the base of observed third (skewness) and fourth (kurtosis)
statistical moments. This is elaborated in some details in Moors [9]. Then, the skewness of a
6 english M. ASADZADEH1,† , E. HASHEMI 1,2 AND A. KOZAKEVICIUS 1,3,‡,∗

(a)

(b)

(−2, 0) x=3

Figure 1: (a) A piecewise linear signal with a singularity at x = 3. (b) The wavelet component.

random variable X is often measured by the standardized third moment γ1 = µ3 /σ 3 , and the
1 PM
fourth standardized moment is defined as γ2 = µ4 /σ 4 , where µi := M i
k=1 (xk − x) , i = 3, 4.
The value of these moments may become arbitrarily large, and therefore hard to interpret.
For a robust measurement, skewness should satisfy the following property, (see [10]):
(i) For any a > 0, and b, γ1 (f ) = γ1 (af + b).
(ii) If f (x) is symmetrically distributed, then γ1 (f (x)) = 0.
(iii) −γ1 (f (x)) = γ1 (−f (x))
(iv) If F and G are cumulative distribution functions of f (x) and g(x), and F <c G, then
γ1 (f (x)) ≤ γ1 (g(x)), where <c is the skewness-ordering among distributions. One aspect in
our study would rely on justification of these properties for the probability density distribution
of sampled data. This is a rather established test in mathematical statistics, see, e.g. [10].
However, there is an equally important statistical measurement quantity: namely the kurtosis,
which either in combination with the skewness, or independently, may be employed for the
analysis of statistical data.
1
Corresponding robustness conditions for kurtosis are not so easy to formulate and therefore
are missing in the literature. Nevertheless, there are other suggestions characterising kurtosis,
e.g. Moors [9] showed that kurtosis is easily interpreted as a measure of dispersion around two
values µ ± σ. For justification of this we shall rely on the following definition:
Definition 2.2.1. For any distribution the kurtosis k is defined as the normalized fourth central
moment (if it exists). Hence, for a random variable X with expectation µ := E(X), variance
σ 2 := V (X) and E(X 4 ) < ∞,

k = E(X − µ)4 /σ 4 ,
i.e., here µ4 is defined as µ4 := E(X − µ)4 .
Introducing a standardized variable Z := (X − µ)/σ yields
k = E(Z 4 ) = V (Z 2 ) + [E(Z 2 )]2 = V (Z 2 ) + 1. (15)
Hence, k may be interpreted as a measure of the dispersion of Z 2 around its expectation 1,
or that of the dispersion Z around ±1. Therefore k measures the dispersion of X around the
two values µ ± σ. One can show that for any symmetric two-point distribution (15), i.e. k,
attains the minimum value 1. This indicates that high kurtosis are possible in two cases: either
 englishMICROCALCIFICATION DETECTION USING WAVELETS 7

concentration of probability mass near µ, or concentration of probability mass in the tails of the
distribution.
An alternative description is based on the following definition.
Definition 2.2.2. For a random variable X the octiles are defined by
P (X < Ei ) ≤ i/8, P (X > Ei ) ≤ 1 − i/8, i = 1, 2, . . . , 7. (16)
For continuous X with distribution function F , the octiles are unique (Ei = F −1 (i/8) for
i = 1, 2, ..., 7), and therefore (16) is simplified as
F (Ei ) = i/8, , i = 1, 2, . . . , 7. (17)
For E6 > E2 a conventional measure proposes an alternative and robust kurtosis measure defined
by
Λ := (E7 − E5 ) + (E3 − E1 )/(E6 − E2 ). (18)
Moors justified this estimator on the ground that the two terms, (E7 − E5) and (E3 − E1), are
large (small) if relatively little (much) probability mass is concentrated in the neighborhood of
E6 and E2 , corresponding to large (small) dispersion around µ ± σ. For further studies we refer
the reader to [9]. Hence, both definitions are indicating that the two approaches yield to the
same characterization of kurtosis as being just a measure of dispersion around µ ± σ.
In our case we specifically deal with a sample of M = 2m0 measurements. Then, for simplicity,
using a somewhat modified notation, the sample skewness S and kurtosis K are given by
1 PM 3 1 PM 4
M k=1 (xk − x) M k=1 (xk − x)
S := γ1 =   3/2
, and K := γ2 =  2 − 3, (19)
1 PM 1 PM 2
M k=1 (xk − x)
2
M (x
k=1 k − x)
respectively, where x is the sample mean.

3. Proposed Methodology
The proposed methodology is based on the main concepts addressed in the previous section
about the application of the Db2 wavelet transform as well as the computation of the two
statistical parameters, skewness and kurtosis.
Since the object of our study is mammogram imaging, the data considered for a cross-section,
will be presented in a matrix Q, based on the two-dimensional discrete Db2 wavelet transform.
The implementation in this case is a straight forward extension of the one-dimensional discrete
Db2 wavelet transform, given by the Cascade Algorithm [8]. In the Algorithm 1 the direct 1D
transform for L levels of decomposition is presented, where the initial vector is denoted by c0 ,
and the output vector by W T 1D(c0 ). The 2D transform is obtained by the application of the
Algorithm 1 to each row of the input matrix and after-wards to each column of the resulting
matrix.
Since the matrix Q has finite dimension (in our applications 29 × 29 ) and the Db2 filters hl , gl
contain 4 values each, the computation of the coefficients cj+1,k and dj+1,k at positions near the
right boundary k = nj − 1, 1 ≤ j ≤ L will require elements from the input vector cj at positions
that will be outside the range of variation of cj . To overcome this issue, a periodic extension
is considered. Values cj,k , whose positions k exceed nj − 1, will then be chosen from values at
the beginning of the vector cj just by considering the modulus operation k̃ = mod(k, nj ). The
main advantage of the choice of Db2 is that this wavelet will cause the smallest interference on
the boundaries because of the periodic extension. For our application this is a desirable feature,
once the microcalcifications can be localized at any position of the mammography.
For our application the number of levels of decomposition L in Algorithm 1 is considered to
be 1, where the wavelet coefficients will be used as a sensor to detect the regions of interest.
We do not need necessarily to reconstruct the image after the analysis. Therefore after one
level of decomposition, each row of Q (the input vector c0 in Algorithm 1) will be initially
8 english M. ASADZADEH1,† , E. HASHEMI 1,2 AND A. KOZAKEVICIUS 1,3,‡,∗

Algorithm 1: WT1D: 1D Direct wavelet transform for J decomposition levels

input : c0 : Initial data vector; n0 = M = 2m0 : vector size;
N = 2 : for Db2, hl and gl : scaling and wavelet filters
output: cJ , dJ , ..., d1 : Set of scaling and wavelet coefficients for J levels
for j ← 0 to J − 1 do
for k ← 0 to nj − 1 do
P −1
cj+1,k = 2N hc
Pi=0−1 i j,2k+i
dj+1,k = 2N
i=0 gi cj,2k+i
nj+1 ← nj /2
return Wavelet decomposition vectors: W T 1D(c0 ) = (cJ , dJ , ..., d1 )

decomposed into two components (c1 , d1 ), producing an intermediate matrix Q̃. We transpose
Q̃ just as an auxiliary operation, and then once again apply Algorithm 1, this time only for one
level decomposition. As the result of the two-dimensional wavelet transform (TW2D) the initial
matrix Q is decomposed into four blocks, Q̃ ˜ = T W 2D(Q) = (cc , dc , cd , dd ) = (C, H, V, D).
1 1 1 1
Here, the block cc1 = C contains the scaling coefficients, and the blocks dc1 = H, cd1 = V and
dd1 = D, are the wavelet coefficients, which capture data variations in the horizontal, vertical
and diagonal directions, respectively.
The next step in our methodology is the computation of the statistical parameters skew-
ness and kurtosis for each block of wavelet coefficients by rows and by columns. Algorithm 2
summarizes these operations and also indicates how the regions of interest are determined.
To specify the regions associated to microcalcification clusters one needs, essentially, to deter-
mine the rows and columns associated to the highest values of kurtosis and skewness computed
for each block H, V, and D. In order to have a criteria to select those highest values, which
are also called significant values, the hard threshold operation T h [7] is considered. The key
ingredient in this operation is the choice of the threshold parameter λ. Since this parameter is
responsible for the entire decision process.
The threshold operation is stated in expression (20) below for a single value s, considering λ
as the threshold value. In Algorithm (2), the threshold operation given by (20) is applied for
each one of the elements of the skewness vectors SrV , SrH and SrD computed based on considering
the row values of the blocks V, H, D. Analogously for the vectors ScV , ScH and ScD computed by
considering the columns of each block. The threshold value λ is computed according to the rule:
λ = 0.80 maxb∈{V,H,D} {maxj {M Sjb }}, taking into account separately the quantities obtained
by rows j = r and those obtained by columns j = c, as indicated in Algorithm (2).
(
s if |s| ≥ λ
T h(s) = (20)
0 if |s| < λ.
After the significant rows and columns were selected by the threshold operation, for each of
the three blocks of wavelet coefficients, for b ∈ V, H, D, the row and column sets Rb and C b are
created. To produce the final selection of significant positions associated to microcalcification
clusters on the original image, the intersections of these sets are considered, i.e., a row or a
column is associated to a microcalcification region if it is significant with respect to the three
blocks of wavelet coefficients, generating the index sets R and C . This criteria is important to
avoid a considerable number of false positive results, as discussed in the next section.
The procedure executed by Algorithm (2) is now repeated substituting the skewness computa-
tions by that of the kurtosis computations of the wavelet coefficients by blocks. For the analysis
based on kurtosis, the same strategy is applied, values are computed by rows and columns of
each block. Again the threshold operation with threshold parameter λ, defined analogously,
 englishMICROCALCIFICATION DETECTION USING WAVELETS 9

Algorithm 2: Regions of interest determination via skewness

input : Qn0 ,n0 Initial data matrix; n0 = M = 2m0
output: R, C : set of significant rows and columns
Selected Regions are neighborhoods of the crossing of elements in R and C
for b in {V, H, D} do
for each row r do
Srb : skewness for row r, block b
M Srb ← maxr {Srb } : Max skewness per block (row computation)
for each column c do
Scb : skewness for column c, block b
M Scb ← maxc {Scb }: Max skewness per block (column computation)
Threshold Operation
Tr ← maxb {M Srb }
Tc ← maxb {M Scb }
S̃rb ← T h(Srb ), λ = 0.80Tr
S̃cb ← T h(Scb ), λ = 0.80Tc
Significant rows and columns per block
Rb = {r : S̃rb 6= 0}
C b = {c : S̃cb 6= 0}
R = b Rb
T
T b
C = bC
return Significant rows and columns R, C
Regions of interest are the crossing sites of significant rows and columns
and their immediate neighboring surroundings

selects the most significant kurtosis values by block. As performed for the skewness parameters,
the rows and columns associated to the significant kurtosis values, finally, would provide us the
index sets R and C, determining the regions of interest. In the next section, numerical experi-
ments considering both versions of the proposed Algorithm (2) are presented. To highlight the
important role played by the association of wavelet coefficients and the skewness and kurtosis
computations, other experiments with some variation in the Algorithm (2) are also designed and
their results are presented and commented as well.

4. Numerical Results and Discussion

The mammographies presented in this section, scanned as raw format with 8-bit gray scale and
512 × 512 pixels size, were obtained from the University of South Florida Digital Mammography
Home Page [14]. This data base also provides annotations from specialists for each image,
indicating the clusters associated to malignant tissues. In order to validate our methodology, 24
images were analyzed, 6 without microcalcification clusters, and 18 containing clusters associated
to malignant tissues.

4.1. Results for the proposed methodology. A mammography with normal tissues is pre-
sented in Figure (2a). Figure (2b) shows a mammography presenting microcalcification clusters
and also the annotation according to [14] (given by the red curve) for the clusters associated to
malignant tissues.
Figure (2c) presents the same mammography, as in Figure (2b), this time however, including
annotations in yellow. The yellow lines were obtained as the result of the Algorithm (2) applied
10 english M. ASADZADEH1,† , E. HASHEMI 1,2 AND A. KOZAKEVICIUS 1,3,‡,∗

for both skweness and kurtosis computations. The intersections of the lines indicate the neigh-
borhoods where microcalcification clusters are more alike. In our simulation two regions were
detected, a cluster associated to the malignant tissue, as well as a second cluster, also associated
to microcalcifications according to [14], but the latter one without malignant tissues.

(a) (b) (c)

(d)

Figure 2: (a)Left breast mammography with normal tissues. (b),(c), (d)Examples with microcalcificatin: Regions of interest
are neighborhoods of the significant rows and columns crossings (Yellow lines). Red curves given by [14]

It is also important to point out that the main purpose of our study is not to determine
whether a cluster is associated to malignant tissues or not. The main goal of our methodology
instead is to provide a support for specialists to pay attention to regions that might not be
observed, but could be of interest. Since in some cases, such regions are not easily detectable,
as can be seen in the examples provided in [14].
Figure (2d) is yet another example of mammography containing microcalcifications cluster.
Again the red curve is the annotation provided by the data bank [14] and the yellow lines are the
results for the analysis obtained by the proposed methodology. In this case, the intersections
of the yellow lines determine the same cluster. In our statistical approach such results are
not considered as false positive results, even-though over determined, the localization of the
microcalcification cluster were done correctly.
In order to collect statistics about the number of correct detections, 24 mammographies
from [14] were analyzed, 6 without any cluster, 18 with annotations for regions with malignant
tissues. For the test group considered normal, the algorithm did not detect any region of interest,
obtaining therefore 100% of correct detection.
For the 18 abnormal cases results are shown in Table (1). The detection results based on the
Algorithm (2), considering skewness, matches to the results obtained when the analysis were
done based on the kurtosis computation, i.e. the same regions were detected when kurtosis was
taken into account. This information is also provided in Table (1), and specified by the wording:
“same region”.
The multiple (more than one) regions identified related to the same microcalcification clusters
that do not characterize false positive results in our statistics, are referred to as over determined
 englishMICROCALCIFICATION DETECTION USING WAVELETS 11

regions. In our experiment the number of false detections were relatively small, three cases
in 18 analyzed mammography with malignant tissues. The column labeled as Annotations in
Table (1) provides the information given by [14], indicating the number of malignant clusters
identified by specialists. We include a remark concerning our over determined detections in the
Annotations columns.

Abnormal Mammographies - Algorithm 2

Image Detection skewness and kurtosis Annotations [14]
skewness kurtosis
1 1 1 same region 1
2 4 4 same region 4
3 1 1 same region 1
4 2 2 same region 1 (over determined)
5 2 2 same region 2
6 1 1 same region 1
7 2 2 same region 2
8 2 2 same region 1 (over determined)
9 (Fig. 2d) 4 4 same region 1 (over determined)
10 4 4 same region 2 (1 false positive)
11 4 4 same region 2 (1 false positive)
12 2 2 same region 2
13 1 1 same region 1
14 6 6 same region 6
15 3 3 same region 1 (1 false positive)
16 (Fig. 2c) 2 2 same region 1 (over determined)
17 2 2 same region 1 (over determined)
18 4 4 same region 1 (over determined)

Table 1: Performance of the proposed detection methodology for abnormal mammography

4.2. Additional simulations. To provide a more precise justification of our methodology,

some additional simulations were performed in order to numerically analyze the efficiency of
the combination of the wavelet transform with the computation of the statistical parameters,
skewness and kurtosis.

4.2.1. Test 1: skewness and kurtosis computed directly from the original input matrix. Now
we apply Algorithm (2) with one major modification, without performing the two-dimensional
wavelet transform. It means that, the values of skewness and kurtosis for rows and columns are
computed directly for the input image data. The same set of 24 mammogram images (6 normal
and 18 with abnormal regions) is considered for the analysis. Table (2) presents the results
obtained for this Test (4.2.1), where now false detections were performed for the set of normal
images. Then the first part of Table (2) shows the results for the three examples with wrong
detections. The detection results for the remaining three normal cases were omitted, since the
detection of the zero cluster was correct.
For the set of 18 images with abnormal clusters the results are presented in the second part
of the Table (2). The column labeled by Identified Malignant regions indicates the number
of matching identified regions according to the number of correct microcalcification regions
specified on the data bank.
The intersections of rows and columns associated to the significant values for skewness and
kurtosis indicate once again the detected regions. Nevertheless, for this experiment the number
12 english M. ASADZADEH1,† , E. HASHEMI 1,2 AND A. KOZAKEVICIUS 1,3,‡,∗

Normal mammography
Image number Detection skewness and kurtosis detections Identified Malignant regions
skewness kurtosis
19 5 5 same 5
20 30 30 same 30
21 12 12 same 12
Abnormal Mammograms
Image number Detection skewness and kurtosis detections Identified Malignant regions
skewness kurtosis
1 2 2 same 0
2 1 1 same 0
3 4 4 same 0
4 6 9 not the same 0
5 12 12 same 0
6 9 9 same 9
7 12 12 same 0
8 16 8 not the same 0
9 12 9 not the same 0
10 15 15 same 0
11 16 16 same 0
12 24 24 same 12 (over determined)
13 5 5 same 5
14 15 15 same 15
15 8 8 same 0
16 30 30 same 0
17 28 28 same 0
18 24 24 same 0

Table 2: Performance using Test 1 on 20 images where 17 of them have abnormal tissues.

of false positives results were much higher than the previous one and the regions eliminated by
the performed analysis, based on skewness and kurtosis, do not match for all analyzed images,
as provided by the Algorithm (2). Another drawback of this test is that: now in some cases,
regions in mammography were assigned to calcification clusters, where in fact no cluster was
present, see for example Figure (3c). The original image is presented in Figure (4a) and the
detection results for the proposed Algorithm (2) are presented in Figure (3b). This experiment
indicates the relevance of the computation of the wavelet coefficients in combination with the
statistical parameters for the correct analysis.

4.2.2. Test 2. This test is basically to investigate whether, or not, the wavelet transform applied
after the computation of skewness and kurtosis, according to the Test (4.2.1), could improve
the analysis or perhaps even avoid the misdirections that were previously obtained in Test
1. Therefore, for the rows and columns, skewness (S r , S c ) and kurtosis ( K r , K c ) were
obtained as described in (4.2.1), and then, the one dimensional wavelet transform is computed
by Algorithm (1), for one decomposition level, generating vectors W T 1D(S r ), W T 1D(S c ),
W T 1D(K r ) and W T 1D(K c ). Then to each one of these vectors the threshold operation (20) is
applied. Once again the threshold value λ is given as in Algorithm (2). We obtain the lines and
columns associated to the most significant values after thresholding (the analysis are preserved
heuristically). Again the obtained information is transferred in a way that they are related to
the original positions of input image.
In Table (3) results for the numerical experiments with 4 abnormal mammographies and 1
normal image are presented, which illustrates the failure of the Test 2 approach.
 englishMICROCALCIFICATION DETECTION USING WAVELETS 13

(a) (b) (c)

Figure 3: (a) Mammography with malignant tissue marked by physicians. (b) Tow regions of interest
selected as being neighborhood of the crossing of significant row and columns obtained by calculating
skewness and kurtosis values of wavelet coefficients. (c) A region selected as being a neighborhood of
the crossing of significant row and columns obtained by the analysis of skewness and kurtosis, computed
directly from the image data indicating no interesting tissue.

Normal mammographies
Image number Detection Skewness and kurtosis detections Identified regions
skewness kurtosis
1 6 6 same 6
Abnormal Mammograms
Image number Detection Skewness and kurtosis detections Identified regions
skewness kurtosis
2 6 6 same 0
3 12 12 same 0
4 10 10 same 0
5 16 8 not the same 0

Table 3: Performance using Test 2 on 5 images where 4 of them have abnormal tissues.

The application of the wavelet transform associated to the threshold operation after the
computation of skewness and kurtosis did not improve any results from the analysis obtained
for Test (4.2.1). Figure (6) also illustrated the detected regions, indicating the failure in the
computation of skewness and kurtosis directly from the original input data.
4.2.3. Test 3. As a final numerical experiment, we would like to confirm the importance of the
application of the wavelet transform in the formulation of the proposed methodology presented
by Algorithm (2). Therefore, in this test we substitute the wavelet transformation by considering
to compute another quantity: the Gradient of the original matrix. We initially compute the
gradient of the image in (x, y)-geometry (GI). The remaining analysis is kept unaffected. Hence,
we follow the Algorithm (2) until the detection part is concluded.
Table (4) shows the performance of Test 3 for 10 mammographies with malignant tissues.
The results presented in Table (4) are for the computation of skewness, computed with respect
to the coefficients obtained by the gradient estimation. In the case where the derivative in x-
direction is computed, five of the presented images were detected as not having any malignant
cluster. And for four other images the number of false positive detections was considerably
high, as presented in Table (4). For the derivative in y-direction, no malignant detection was
performed for three images and for the remaining six images the number of false detections was
again considerably high. The results obtained for the computation of kurtosis were omitted,
14 english M. ASADZADEH1,† , E. HASHEMI 1,2 AND A. KOZAKEVICIUS 1,3,‡,∗

(a) Image 2 in Table 3 (b) Image 3 in Table 3

Figure 4: Test 2 selected regions, where detected clusters do not match with interested micro-
calcification clusters.

Abnormal Mammograms
Image number Detection Detected region by skewness of Gx Detected region by skewness of G
skewness of Gx skewness of Gy
1 4 4 4 4
2 9 8 2 1
3 25 24 0 8
4 40 16 0 8
5 56 24 20 0
6 30 12 0 3
7 36 21 15 9
8 6 8 0 0
9 28 20 0 0
10 42 20 8 3

Table 4: Performance using Test 3 for 10 images with abnormal regions.

since the quality of the detection considering the modification proposed in this test was neither
satisfactory.

(a) Image 10 in Table 4 (b) Image 3 in Table 4

Figure 5: Regions selected as being neighborhood of the crossing of significant rows and columns
obtained by the analysis of derivatives in y- direction of skewness and kurtosis computed from the
image data. (a) indicates no interesting tissue and (b) indicates the over detected of malignant
tissue and plenty of false detections.
 englishMICROCALCIFICATION DETECTION USING WAVELETS 15

(a) Image 9 in Table 4 (b) Image 5 in Table 4

Figure 6: Regions selected as being neighborhood of the crossing of significant row and columns
obtained by the analysis of derivative in x- direction of skewness and kurtosis computed from
the image data, where (a) indicates no interesting tissue and (b) indicates the over detected of
malignant tissue with plenty of false detections.

For the association of the wavelet coefficients with the computation of the third and fourth
statistical moments (skewness and kurtosis), the ordering in the computations does matter for
the entire algorithm performance; and the computation of these two statistical measurements
based on wavelet coefficients is more efficient for the determination of malignant regions in
mammographies than the computation of skewness ad kurtosis directly from the image or from
another transformation.

5. Conclusion
This work presents an approach to detect microcalcifications of mammography based on the
two dimensional Daubechies wavelet transform and the computation of two statistical parame-
ters, skewness and kurtosis. The potential of the three wavelet subbands of the mammography
decomposition is explored to identify regions associated to microcalcifications clusters on the
original image.
The importance of the use of the decimated wavelet transform with two vanishing moments
(Db2) is argued analytically, as well as numerically. The performance of the proposed methodol-
ogy, involving the analysis of the computed statistical parameters, was tested by reliable results
for the detection of regions associated to microcalcifications. The set of 24 images used for the
numerical tests were obtained from the American Cancer Society.
Also different associations were verified, though they were in poor quality and were not
considered to be satisfactory. These variations were explained with Tests 1, 2 and 3. The first
was the computation of skewness and kurtosis directly from the input matrix, the second was
the application of the one-dimensional wavelet transform after the computation of the statistical
parameters, and for the lasr test a different transformation at the initial stage of the proposed
Algorithm (2) was considered. None of the modifications had produced results comparable to
the performance presented by our Algorithm (2), what provides here an evidence of the relevance
of the proposed algorithm to perform the analysis.
According to the numerical evidences presented by the tables and illustrated by the images,
the methodology proposed and discussed in this paper is an easy to implement and effective
alternative for the analysis of mammography, with the goal of detecting regions associated to
microcalcification clusters.

References
[1] M. Asadzadeh, Lecture Notes in Wavelets with Fourier Analysis, Electronic version, Chalmers (2010).
[2] Jöran Bergh, Fredrik Ekstedt, Martin Lindberg, Wavelets, Studentlitterature, Lund (1999).
16 english M. ASADZADEH1,† , E. HASHEMI 1,2 AND A. KOZAKEVICIUS 1,3,‡,∗

[3] Albert Boggess and Francis J. Narcowich, A first Course in Wavelets with Fourier Analysis , Prentic Hall,
(2001).
[4] M. Garcia, A. Jemal, Global Cancer Facts and Figures 2011, Atlanta, GA: American Cancer Society (2011).
[5] Ingrid Daubechies, Ten Lectures on Wavelets (1992).
[6] Lokenath, Debnath, Wavelets Transforms and their Applications, Birkhäuser, (2002).
[7] David L. Donoho, De-noising by Soft-Thresholding IEEE Transactions on Information Theory, Vol. 41 No.
3, May 1995.
[8] Mallat, S. A Wavelet tour of signal Proccesing, Academic Press, San Diego, 1998.
[9] Moors, J. J. A. A Quantitative Alternative for Kurtosis, J. the Rotal Stat. Soc. Ser. D (The Statistician) 37,
pp. 25-32 (1988).
[10] H. Oja, On location, scale, skewness and kurtosis of Univariate Distributions, Scand, J. Stat. pp 1-15 (1980).
[11] K.Prabhu Shetty, V. R. Udupi and K. Saptalakar, Wavelet Based Microcalcification Detection on Mam-
mographic Images , Intenational Journal of Computer Science and Network Security, VOL. 9 No. 7, July
2009.
[12] Ted C. Wang , Nicolaos B. Karayiannis, Detection of Microcalcifications in Digital Mammograms Using
Wavelets, IEEE Transactions on medical imaging, VOL. 17, No.4 (1998).
[13] http://www.cancer.org/Cancer/BreastCancer/, American Cancer Society
[14] http://marathon.csee.usf.edu/Mammography/Database.html, University of South Florida Digital Mammog-
raphy Home Page