A comparison of some gradient threshold

estimators for the Nonlinear Diffusion of

Perona-Malik. A novel proposal to improve edges
detection in mammograms.

Reinaldo Barrera Travieso1 , Angela M. León-Mecías2[0000−0001−7212−5783] , José

A. Mesejo-Chiong2[0000−0001−7230−0043] , and Richard M. Méndez-Castillo3
1
Instituto de Geofísica y Astronomía, La Habana, Cuba
2
Facultad de Matemática y Computación Universidad de la Habana, San Lázaro y
L, Habana 4, CP-10400, Cuba
{angela,mesejo}@matcom.uh.cu
3
Union de Informáticos de Cuba, La Habana, Cuba

Abstract. The aim of this contribution is to make an efficient smooth-

ing algorithm for mammography images that preserves edges and pro-
vides valuable information for any segmentation process. The non-linear
anisotropic diffusion (AD) model of Perona–Malik is considered to en-
hance the edges in the process of diffusion through a variable diffusion
coefficient. However, the diffusion coefficient is very sensitive to the so-
called contrast or gradient threshold parameter. This work proposes an
algorithm to estimate the threshold of the gradient in the diffusion coeffi-
cient, so that it can be applied to mammography images. It is compared
with two others way of estimating the gradient threshold. The experimen-
tal results show that the quality of the edge detection process improves
when AD is used. When the proposed algorithm (AD-KMLS2) is used,
results are better for 45% of the images in terms of two quality measures:
the so-called Pratt’s figure of merit and the Root Mean Square Error.
Although, the difference is of the order least equal 5 × 10−5 when it does
not achieve the best results. The anisotropic diffusion smoothing is done
by region, using the superpixel segmentation technique SLIC.

Keywords: Perona-Malik model · KMLS2-gradient threshold estimator

· mammography images.

1 Introduction

Mammograms are images with low contrast and a complicated structured back-
ground and generally contaminated by some noise. At the same time, mammo-
grams are very important non-invasive tests in the early detection of breast can-
cer, a disease that according to the World Health Organization (WHO), reaches
more than 2 billions of women every year, and is considered the second leading
cause of death from cancer among women, after lung cancer [1]. In 2018, 3 748
2 R. Travieso et al.

women were fatal victims of the disease in Cuba, which represented the 40% of
cases diagnosed that year4 . Then it is important to have algorithms that im-
prove the quality of mammograms so that specialists can make more accurate
diagnoses. Image smoothing is one of the most common task in image process-
ing, with the aim to improves image quality by some criteria, it means to obtain
noise free sharpness images ready for a forthcoming analysis. In this work we
are interested in image smoothing algorithms that keeps edges, edges from detail
existing at fine scales and edges from the boundaries of large objects existing at
large scales; which is particularly important as previous step for edge detection
and image segmentation.
Among the image smoothing PDE-based methods, Perona and Malik [2] pro-
posed for the first time a non linear diffusion based algorithm that overcame the
disadvantages of the linear diffusion such as blurring or dislocates the edges of
the images. Nonlinear anisotropic diffusion (AD) is a filtering tool that preserve
the position of such edges through the image smoothing process. In the seminal
paper [2], they proposed a non linear PDE with variable diffusion coefficient to
model the diffusion process. The diffusion coefficient c is chosen as an appropriate
function of the gradient of the brightness function c(x, y, t) = g(k∇I(x, y, t)k),
that encourages intra-region over inter-region smoothing. To achieve a better
performance in edge preserving, different expressions for the diffusion coefficient
have been proposed, see for instance Table 1 in [3]. All the given expressions for
this coefficient depend on a parameter k, called contrast parameter or thresh-
old for the magnitude of the gradient of the brightness with great influence on
edge preserving. There are few works that focus on estimating the contrast pa-
rameter k. The simplest is to use a constant k, with the difficulty that it does
not take into account smoothing performed on the previous scales and applies a
smoothing with equal intensity in each iteration. In order to decrease k for each
iteration and thus maintain the condition k∇Ik > k to preserve the edges, there
are some alternative methods as can be consult in [3–5]. Particularly interesting
for us is the proposal in [3], where a new method called Partition and Adjust-
ment (KMLS) to estimate k for gray scale images is presented. This method
combine the K-means algorithm and a least square fit to approximate the diffu-
sion coefficient. However, it does not behave well for all types of images, among
which we can mention mammograms, see [4].
This work aims to apply AD of Perona-Malik by regions in mammograpy
images to remove noise and heighlight edges, to compare different ways of esti-
mating the threshold of the gradient magnitud in the diffusion coefficient, and
its influence on edge detection. For this, a new estimator for the threshold of the
magnitud of the gradient is proposed. It is a modification of the one presented
in [3]. A SLIC technique is used to partitioned the image.

4
https://adncuba.com/noticias-de-cuba/actualidad/minsap-da-estadisticas-del-
cancerde-mama-fallecio-el-40-de-los-casos
 Title Suppressed Due to Excessive Length 3

2 Image smoothing by diffusion

Diffusion is known as a physical process that equilibrates concentration differ-

ences without creating or destroying mass. In image processing we may identify
the concentration with the intensity of the gray value at a certain location. If
the diffusion tensor is constant over the whole image domain, one speak of ho-
mogeneous or isotropic diffusion. Isotropy in image diffusion is interpreted such
as every pixel is treated the same way, independently from its characteristics
like the gradient. This brings some drawbacks: for instance, applying Gaussian
2 2 2
filter Gσ (x, y) = (1/2πσ 2 )e−(x +y )/2σ with successively larger σ values, Gaus-
sian smoothing does not only reduce noise but also blurs important features
such as edges and thus makes them harder to identify, [6]. The resulting Gaus-
sian scale-space of images may also be viewed as the evolving solution of a PDE √
(∂I(x, y, t))∂t = c∆I(x, y, t) where the time-scale relationship is given by σ = t
[6]. Due to the valuable information we can get from edges in an image, Perona
and Malik [2] proposed a non linear PDE model, to avoid image blurring and
localization difficulties of linear diffusion filtering, given by

∂I(x, y, t)
= div(c(x, y, t) · ∇I(x, y, t)), in Ω × (0, T ) (1)
∂t
c(x, y, t) · ∇I(x, y, t) · N = 0, in ∂Ω × (0, T ) (2)
I(x, y, 0) = I0 (x, y), in Ω, (3)

where Ω ⊂ R2 , I(x, y, 0) is the original image intensity function and I(x, y, t)

is the smoothed image intensity function in time t, and N is the normal direc-
tion. In what follows I(x, y, t) = It (x, y). The diffusion coefficient c is chosen as
an appropriated function of the gradient of the brightness function c(x, y, t) =
g(kIt (x, y)k), to control the diffusion near the edges. In [2] two expressions were
−s2
considered: g1 (s) = 1
2 and g2 (s) = e( 2k2 ) . Although, the Perona–Malik model
1+ s 2
k
is called of AD, in the field of differential equations this term is used when the
diffusion coefficient c is a tensor, [6]. As Perona and Malik remark, their model
is ill posed, [7]. However, the authors in [8] proved that a standard spatial fi-
nite difference discretization is enough to turn the Perona-Malik process into a
well-posed system of nonlinear ordinary differential equations.

2.1 The gradient threshold in the diffusion coefficient

Many interesting variations of the Perona-Malik model have been proposed, most
of them related to the regularization of the model considering the introduction
of new expressions for the diffusion coefficient, see for instance [6, 9]. This coeffi-
cient characterizes the diffusion process and the goal is to carry on the smoothing
inside regions preferentially over the smoothing along edges. The diffusion coef-
ficient also works as a time stopping function and all known expressions for it
depend on a parameter k in such a way that in those region of the image where
k∇Ik > k , the smoothing effect is weaker and therefore the edges are preserved
4 R. Travieso et al.

and where k∇Ik < k, the smoothing effect is stronger. It means k is a contrast
parameter or a threshold for the magnitude of the brightness gradient with great
influence in edge preserving. With the same number of iterations the result of
diffusion process for different values of k is also different. If we increase the value
of k, the diffusion process blurs the image, removing the details. During diffusion
the image is smoothed which means, the value of the gradient decreases, then the
parameter k that controls the diffusion rate must be readjusted in each iteration,
so that edges remains preserved. It is evident that the diffusion coefficient is very
sensitive to the choice of the parameter k. At the same time, k has to be chosen
considering not only global but local changes of the brightness gradient. Due
to, a diffusion process by regions is proposed in this work, dividing the image
using the SLIC segmentation algorithm, [10]. Then k should also depends on
time as was proposed in [3], where a method based on K-means clustering and
least square fitting (KMLS) to estimate k was introduced. Although the method
in [3] shows good results for Natural Images, its performance for mammograms
is very poor, even sometimes it doesn’t work, [4]. The goal of the present paper
is to address this topic by evaluating the quality of the edges for mammogram
images, and comparing it with other ways of estimate k.

3 Estimating the gradient threshold for mammograms

In the literature review there are some ways to estimate the gradient threshold
k, but they they don’t work in the same way for all kind of images. In this
article, some of the them were selected with the aim to verify their effectiveness
for mammograms and to compare them with our proposal.
In order to decrease the value of k for each iteration and thus maintain
the condition k∇Ik > κ to preserve the edges, Francesco Voci [5] proposed
two alternative methods for setting k. A first one known as p-norm, where the
σ·kIt kp
threshold value is calculated as: κ = n·m , with It representing the image
I(x, y) in the t iteration. Because of this value is too high, adaptive scaling of
an image with n rows and m columns will be used. σ is a constant proportional
to the average intensity of the image. A second method presented in [5] is based
on a morphological approach. If the “open–close” difference is taken into account
as the quantization parameter, then κ must be at least equal to the averaged
noise amplitude value. Given an image It with n rows and m columns, and
given a structuring element st (usually 3 × 3 or 5 × 5), we can estimate κ as
follows: κ = (x,y)∈It (It (x,y)◦st) − (x,y)∈It (It (x,y)•st)
P P
n·m n·m , where the symbols “◦”
and “•” represent the open and close operation, respectively. Another methods
to estimate k is the noise estimator (Noise), initially described by [11], and later
used by [2]. In this method a histogram of the absolute values of the gradient in
the whole image is calculated and k is set equal to 90% of its integral value at
each iteration.
In [3] an algorithm called KMLS is proposed. It considers updating the
threshold of the gradient of the intensity in each iteration. Applying a clustering
K-Means, the set P of pixels is partitioned into three subsets: P1 , the subset of
 Title Suppressed Due to Excessive Length 5

pixels that does not belong to the edges of the image, P3 subset of pixels belong-
ing to the edges of the image and P2 = P \ (P1 ∪ P3 ), subset of pixels called fuzzy
set because its elements can be weak edges or cannot be edges. From those sets,
the points used to approximate the contrast parameter k are selected, through a
Least Square curve fitting of the curve defined by the diffusion coefficient g2 (s)
given above. Let i21 and i31 be the minimal values of the gradients of the pixels
that belong to P2 and P3 respectively. To control the smoothing strength two
thresholds are defined: wep and sep of weak and strong edges preservation re-
spectively. These thresholds are in practice the values of the diffusion coefficient
g2 (s) that we want to make correspond to the values of the gradient of inten-
sities i21 and i31 . The pixels in P3 (strong edges) are smoothed with strength
less than sep and the pixels of weak edges are smoothed with strength less than
wep; therefore sep should be taken as close as possible to zero. On the other
hand, the threshold wep should be selected taking information from the image,
nevertheless this seems to be a pretty hard work, and requires experimenting in
a specific kind of images, to get some values from statistics. Having the points
(i21 , wep) and (i31 , sep), the contrast parameter k will be obtained using Least
Square fitting, [3]. We focused our attention in this estimator to applied it in
mammograms smoothing, but it was found that it works well for natural images
in scale of gray, but when we tried to apply it in mammogram images, it has
some drawbacks, [4].

3.1 Problems with the gradient threshold estimator KMLS for

mammogram images

The KMLS algorithm in [3] shows us the process of partitioning according to

K-Means, where is evident the existence of a minimum and a maximum values
of the intensity gradient of the images but you may also find values greater than
the minimum, and values less than the maximum. Those intermediate values are
represented in the KMLS algorithm, as the mean of the two extreme values. But
such values need not exist, and this could guide to a problematic situation when
the cluster with this centroid remains empty.
Usually, images have contrast enough and there is lots of variation in in-
tensities of the pixels. However, this is not the case with mammograms, which
are low-contrast images due to the low doses of radiation to which patient are
subjected. On the other hand, if the image is divided into regions before the
smoothing, as is our goal, it may happen that every pixel in it have similar gra-
dient values; either because the segment is a border of the same zone and the
values are equal and positive, or because this region is entirely of the same color
and the gradient is zero. Under such conditions the performance of the method
KMLS is very poor or simply does not work. The low contrast problem led to the
idea of introducing the image segmentation before running any diffusion process,
trying to locally smooth the image, to reinforce small edges. To this purpose su-
perpixel based segmentation will be used, [4], and some modifications to KMLS
algorithm will be proposed.
6 R. Travieso et al.

3.2 KMLS2, a modified version of KMLS for AD smoothing of

mammogram images

To overcome the drawbacks of KMLS [3], we propose some modifications, al-

though preserving the original ideas of Partitioning and Fitting. The new ver-
sion will be called KMLS2. To get the partition of the set of pixels of the image,
K-Means will be executed for M = 2, in order to solve the drawback of having
the P2 cluster, defined in KMLS, empty. Then, there are only two clusters of
the intensity gradient set P , denoted as G1 and G2 . It is important to note that
in this case every cluster will have at least one values that belongs to it. The
way to select the magnitudes of the gradient to perform the least square fitting
for the function g2 (s) are also modified. The gradient magnitude values i21 and
i31 are selected as the value of the final centroid of K-Means for the set G1 and
G2 respectively. As for KMLS [3], to control the intensity of the smoothing pro-
cess over the different subsets, two thresholds are defined: weak edge preserving
(wep) and strong edge preserving (sep). These two thresholds are values that the
diffusion coefficient will take for two values of the magnitude of the gradient. To
illustrate the procedure we estimate the contrast parameter for the coefficient
expression g2 as in [3].
Step1: Apply K-Means to get the partition in two clusters, {G1 , G2 } of the set
P of pixels of the image, where:

– Initial Means µ1 = min{k∇Ik(x,y) : (x, y) ∈ P } f or G1 , µ2 = max{k∇Ik(x,y) :

(x, y) ∈ P } f or G2 ,
– Operations

(i) Distance: d((x, y), µi ) = |k∇Ik(x,y) − µi |,i = 1, 2.

(ii) Add:a((x1 , y1 ), (x2 , y2 )) = k∇Ik(x1 ,y1 ) + k∇Ik(x2 ,y2 ) .
 k∇It (x,y)k 2
−
Step2: Make the Least Squares fitting of g2 (k∇It (x, y)k) = e k
. By
linearization we obtain
n
" !#2
X −x2j
S = k ln f − Gk2 = min ln f (xj ) − (4)
k
j=1
k2

where G = ln g2 , xj = k∇It (xj , yj )k, n = 2, having the points (i21 , wep) and
(i31 , sep). The gradient magnitude values i21 and i31 are selected as the value of
the final centroid of K-Means for the set G1 and G2 respectively. Then, solving
the equation
dS
= 0, (5)
dk
we obtain the following expression for the contrast parameter k
s
−(i421 + i431 )
k= . (6)
ln(wep)i221 + ln(sep)i231
 Title Suppressed Due to Excessive Length 7

Since the diffusion process will be carried out by regions, it may happen that one
of these regions is also of very low contrast, that is, the values of the intensity
gradient are almost identical. In that case the initial centroids for the clusters in
K-Means will be equals (µ1 = µ2 ). Then, the fitting phase will be done using only
0 0 0 0
one point, (i0 , g2 (i0 )), with i0 = µ1 = µ2 . To choose g2 (i0 ) we take M axG and
M inG as the biggest and lowest gradient of the whole image, respectively. Then,
0 0
if: M axG − i0 < i0 − M inG holds, the sep threshold is used to fit g2 (s) making a
weak smoothing process. On the contrary wep will be used for a strong smoothing
process. In the case i0 0 = 0 then k = 0, which brings a complications because
it undefines the function g2 (s). However, as i0 0 = 0 the coefficient c(x, y, t) = 1
and we do not need to fit the curve. This is equivalent to an isotropic diffusion in
a homogeneous image (without contrast or edges), which is fine, because there
are not characteristics to preserve.

4 Experimental environment and results

In this section, we use the AD equation of Perona–Malik with different ways to

estimated the contrast parameter, using the diffusion coefficient given by g2 to
smooth the mammogram images, followed by the classical edge detector Scharr
[12] to obtain an edge image. This edge detector was chosen as a result of pre-
vious experimentation, with 20 images from the Inbreast database [13] as you
can see in Table 1. The Perona-Malik equation (1) is solved applying the same
finite difference scheme as in [2]. The aim is to show how the previous filtering
of the mammography image using AD, greatly improves the subsequent edge
images. To choose the contrast parameter estimator with better performance
is also a goal. Because ground truth edge images of the entire breast are not
available, but only of the lesions known as masses, the border images that are
compared are those mammograms where this anomaly appears, using the IN-
breast data base [13]. To avoid that each expert may have different criteria
and preferences and consequently, the reviews given by two experts can differ
substantially, we used two quantitative performance error measures based on
distance: i) the well-known Root Mean Square Error (RMSE)[14] and, ii) an
alternative formulation of Pratt’s figure of merit (PFoM*) [3]. PFoM* is defined
as P F oM ∗ (RI, GT ) = 1 − P F oM (RI, GT ), see [15] for PFoM. For the experi-
mentation with our proposal KMLS2, to estimate the contrast parameter we use
the fixed combination of values sep = 0.05 and wep = 0.3. In Fig.1 and Fig.2, we
can see the evolution of the objective performance measures as a function of the
values of sep and wep respectively for the Scharr edge detector. Figure 1 shows,
for the fixed value wep = 0.3, the Pratt and RMSE measures as a function of
sep values, for 10 of the images in the database Inbreast. As can be seen, for
some fixed images, the variation of measures values is very soft and its differ-
ences for different values of sep are very small. However for others fixed images
the variation of measures values can be very different in general, and between
different images it also changes quite a bit. The shape of the graphs of Pratt
and RMSE measures in function of sep values, for a fixed value of wep, are very
8 R. Travieso et al.

Table 1. Table captions should be placed above the tables.

ImageID RMSE PFoM

Scharr Sobel Farid Prewitt Scharr Sobel Farid Prewitt
53586896 0.708959 0.709008 0.709383 0.709022 0.052801 0.052802 0.052807 0.052846
50997515 0.688559 0.688655 0.688647 0.688703 0.049961 0.049963 0.049962 0.049964
53582656 0.482549 0.482936 0.485803 0.483191 0.025148 0.025153 0.025188 0.025158
53580804 0.697287 0.697398 0.697531 0.697475 0.051139 0.051166 0.051140 0.051144
50999008 0.737287 0.737288 0.738493 0.737292 0.056780 0.056780 0.056809 0.056781
51049053 0.526704 0.526713 0.527303 0.526732 0.029874 0.029874 0.029885 0.029874
51048738 0.661380 0.661388 0.662508 0.661396 0.046264 0.046264 0.046289 0.046264
51049107 0.494382 0.494385 0.495048 0.494387 0.026417 0.026417 0.026430 0.026417
50999432 0.535592 0.535651 0.535595 0.535690 0.030849 0.030850 0.030850 0.030850
53581406 0.789987 0.789992 0.792180 0.790017 0.064554 0.064554 0.064618 0.064554
50999459 0.772492 0.772626 0.772776 0.772715 0.061993 0.061996 0.062001 0.061998
53582683 0.561914 0.562371 0.565213 0.562668 0.033847 0.033853 0.033890 0.033857
53587014 0.615225 0.615265 0.615594 0.615284 0.040289 0.040290 0.040294 0.040291
53586960 0.608030 0.608136 0.608332 0.608208 0.039406 0.039408 0.039415 0.039410
53587663 0.420495 0.420525 0.421585 0.420543 0.019256 0.019257 0.019283 0.019330
51048765 0.662528 0.662536 0.663370 0.662543 0.046399 0.046399 0.046418 0.046399
53580858 0.774775 0.774901 0.774947 0.774969 0.062357 0.062360 0.062364 0.062362
53586869 0.642760 0.642772 0.643348 0.642762 0.043815 0.043816 0.043825 0.043816
53581460 0.687165 0.687180 0.688590 0.687204 0.049734 0.049734 0.049769 0.049734
50998981 0.715995 0.716027 0.717161 0.716042 0.053748 0.053749 0.053775 0.053749

similar. Nevertheless, the value sep = 0.05 provides a good value of the measure
for 9 of the 10 images. In Fig.2 we show the objective performance measures as
a function of the values of wep, for a fixed value sep = 0.05. Again, the graphs
of Pratt and RMSE measures are very similar. Reasoning in the same way, we
choose wep = 0.3.

4.1 Results and discussion

Using the same 20 images as before from the Inbreast database [13], the crop
function was applied to each image to reduce unwanted external areas to the
breast. For these images, the pixel meshes and the masks of the ground truth
of the lesions were extracted. In a previous step the images were divided into
regions using the SLIC clustering algorithm, with 1, 6, 12, 20, 30 and 100 clusters,
with the best results being for 6 clusters, used for the experiments that follow.
Three types of Anisotropic Diffusion (AD) smoothing were performed, that differ
in the way of estimating the threshold parameter k of the gradient in the diffu-
sion coefficient, namely: using pnorm (AD-pnorm) [5], using the noise estimator
(AD-Noise) [11], and using the proposed algorithm KMLS2 (AD-KMLS2), a
modification of KMLS [3]. The showed results correspond to 10 iterations. After
that, the classic edge detector Scharr is applied to each smoothed image. The
performance of the algorithms is first visually measured comparing the edge
images obtained for each filtered image, as shown in Fig. 3 and Fig. 4. The
 Title Suppressed Due to Excessive Length 9

Fig. 1. Evolution of the objective performance measures as a function of sep values,

for a fixed value wep = 0.3 and the Scharr edge detector. Left: PFoM* measure. Right:
RMSE measure.

Fig. 2. Evolution of the objective performance measures as a function of wep values,

for a fixed value sep = 0.05 and the Scharr edge detector. Left: PFoM* measure. Right:
RMSE measure.
10 R. Travieso et al.

quality of the edge images using AD is always better than when pre-smoothing
is not applied. In both figures, visually the difference between AD-KMLS and
AD-Noise is the same, that in the later the edges are thicker. However using the
objective quality measures, for the image ID53586896 AD-KMLS has a better
result, (Fig. 5) and for the image ID53587663, the better result is for AD-Noise.
An interesting observation is that the visual results are better when the value of
the objective measures is close to 0.06 for PFoM* and close to 0.7 for RMSE.

Fig. 3. Mammogram ID53586896 (a) Mask of the anomaly ground truth edge image,
edge images: (b) without smoothing (c) and(d) after AD-KMLS2 and AD-Noise re-
spectively.

Fig. 4. Mammogram ID53587663 (a) Mask of the anomaly ground truth edge image,
edge images: (b) without smoothing (c) and(d) after AD-KMLS2 and AD-Noise re-
spectively.

If we apply our algorithm (AD-KMLS2) we obtain better results for 45% of

the images in terms of the two quality measures. Although, when the results with
AD-KMLS are not the best, the difference is of the order least equal 5 × 10−5
compared to AD-Noise, which is the other estimator with good results, see Fig. 6.
 Title Suppressed Due to Excessive Length 11

Fig. 5. Objective performance measure RMSE(left), PFoM*(right), applied to the edge

image of ID53586896. Smoothing by AD-KMLS2 (blau), AD-pnorm (red) and AD-
Noise (green).

Fig. 6. Objective performance measure RMSE(left), PFoM*(right), applied to the edge

image of ID53587663. Smoothing by AD-KMLS2 (blau), AD-pnorm (red) and AD-
Noise(green).

5 Conclusions

In this contribution, a methodology to estimate the gradient threshold in the

Perona–Malik AD filter for mammograms is presented. It is a modification to
that introduced in [3], which is based on a partition of the pixel set of the
image, according to its gradient magnitude and combined with a least square
fitting for the diffusion coefficient. The AD is done by regions, using SLIC as
segmentation algorithm. At each integration step, a new contrast parameter is
obtained for each region of the image. The experiments, using 20 images with
mass-like anomalies from the INbreast database show that the quality of the
edge images is superior when a previous AD smoothing is applied. The obtained
12 R. Travieso et al.

edge images were evaluated using the PFoM* and the RMSE error measures. The
proposed method AD-KMLS2 achieved better results for 45% of the images in
terms of both quality measures. For the rest of the images the AD-Noise worked
better, although the difference according to the measurements is of the order
least equal 5 × 10−5 and visually the differences are barely perceptible. The best
results are reached by Scharr edge detector.

Acknowledgements We thank the Research Project "Métodos numéricos para

problemas en múltiples escalas", Ciencias Básicas y Naturales, CITMA, Cuba.

References
1. Shelley McGuire. World cancer report 2014. geneva, switzerland: World health or-
ganization, international agency for research on cancer, who press, (2016)
2. Perona Pietro and Malik Jitendra: Scale space anisotropic and edge detection using
anisotropic diffusion. IEEE Transactions on Pattern Analysis and Machine Intelli-
gence 12(7), 629—639 (1990)
3. Borroto-Fernández M., González-Hidalgo M., León-Mecías: New estimation method
of the contrast parameter for the Perona–Malik diffusion equation. Computer Meth-
ods in Biomechanics and Biomedical Engineering: Imaging & Visualization 4(3-4),
238–252 (2014)
4. Hidalgo-Gato, Elizabeth: Estimación del pámetro de contraste para el suavizado
por Difusión Anisotrópica aplicado por regiones. Tesis de Licenciatura, Universidad
de La Habana, Cuba (2015)
5. Voci F, Eiho S, Sugimoto N, Sekibuchi H.: Estimating the gradient in the Per-
ona–Malik equation. IEEE Signal Process Mag. 21(3), 39—65 (2004)
6. Weickert, Joachim: Anisotropic Diffusion in Image Processing. ECMI, Series. Teub-
ner, Stuttgart (2008)
7. Catté F., Lions P., Morel J. and Coll T.: Image selective smoothing and edge de-
tection by nonlinear diffusion. SIAM J. Numer. Anal. 29, 182–193 (1992)
8. Weickert, Joachim and Benhamouda, B.: A semidiscrete nonlinear scale-space theory
and its relation to the Perona-Malik paradox. Springer, Wien (1997)
9. Gilboa G, Sochen N, Zeevi YY.: Forward-and-backward diffusion processes for adap-
tive image enhancement and denoising. IEEE Trans Image Process. 11(7), 689—703
(2002)
10. Achanta, Radhakrishna, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua
and Sabine Susstrunk: SLIC superpixels compared to state-of-the-art superpixel
methods. Pattern Analysis and Machine Intelligence, IEEE Transactions. 34(11),
2274—2282 (2012)
11. Canny, J.: A Computational Approach To Edge Detection. IEEE Trans. Pattern
Analysis and Machine Intelligence 8, 679–714 (1986)
12. Scharr, H.: Optimal Operators in Digital Image Processing. PhD Thesis, University
of Heidelberg (2000)
13. Moreira, I., Amaral, I., Domingues I. et. al : INbreast: Toward a Full-field Digital
Mammographic Database. (2011)
14. Wikipedia, Mayo 2015. http://en.wikipedia.org/wiki/ Root-mean-
square_deviation
15. Abdou I., Pratt W.: Quantitative design and evaluation of enhance-
ment/thresholding edge detectors. Proc IEEE. 67(5), 753—763 (1979)