close

Вход

Забыли?

вход по аккаунту

?

s10916-017-0839-8

код для вставкиСкачать
J Med Syst (2017) 41:190
https://doi.org/10.1007/s10916-017-0839-8
TRANSACTIONAL PROCESSING SYSTEMS
Detection and Segmentation of Pectoral Muscle
on MLO-View Mammogram Using Enhancement Filter
P. S. Vikhe1 · V. R. Thool2
Received: 22 November 2016 / Accepted: 11 October 2017
© Springer Science+Business Media, LLC 2017
Abstract The presence of predominant density region of
the pectoral muscle in Medio-Lateral Oblique (MLO) view
of the mammograms can affect or bias the results of mammograms processing for breast cancer detection using intensity based methods. Therefore, to improve the diagnostic
performance of breast cancer detection using computeraided system, identification and segmentation of pectoral
muscle is an important task. This paper presents, an intensity based approach to identify the pectoral region in mammograms. In the presented approach enhancement mask and
threshold technique is used to enhance and select the pectoral region and boundary points respectively, to find the
boundary of pectoral muscle. Then curve fitting by Least
Square Error (LSE) method is used to refine the rough
initial boundaries. The proposed approach was applied on
320 mammograms from mini-Mammographic Image Analysis Society (mini-MIAS) database of 322 mammograms,
with acceptable rate of 96.56% from radiologist experts.
The performance evaluation for pectoral muscle segmentation, based on Hausdorff distance (Hd ), False Positive (FP)
and False Negative (FN) rate, shows the usefulness and
effectiveness of the proposed approach.
This article is part of the Topical Collection on Transactional
Processing Systems
P. S. Vikhe
pratapvikhe@gmail.com
V. R. Thool
vrthool@yahoo.com
1
Pravara Rural Engineering College, Loni (MS), India
2
S.G. G. S. I of E & T, Nanded (MS), India
Keywords Mammograms · Pectoral muscle ·
Enhancement mask · Segmentation · Detection
Introduction
The breast cancer is most recurrent health problem diagnosed among womens, in both developing and developed
countries, that leads to cause death. In the United States
(US), one out of eight womens will develop breast cancer
at some stage during her life time according to National
Cancer Institute [1]. In 2004, according to World Health
Organization (WHO), cancer accounted 13% of all deaths
in the world [2]. Breast cancer cases and deaths estimated
by American Cancer Society (ACS), yearwise of US are
shown in Fig. 1 [3]. The Fig. 1a–c presents the approximate
new breast cancer cases with dark black bar and approximate deaths with light bar yearwise for both sex (male and
female), female and male respectively. The radiologist use
mammography widely as a diagnostic and screening tool,
to detect breast cancer at early stage. It is the most reliable technique for early detection of breast cancer, reducing
mortality rates up to 25%. Screening mammography is not
easy task for radiologists, 10%-30% of lesions are missed
during routine screening [4–6]. The double screening of
mammograms can increase the accuracy, but expertise limitations and intra-observer variability limit its use due to
huge number of mammograms. The accuracy and speed can
be improved using computerized mammographic analysis.
Computerized analysis can reduce the work of radiologists
and discrepancies due to inter-observer. But, inclusion of
predominant density region of pectoral muscle, in most
MLO views of mammograms, can bias or affect the results
of cancer detection during computerized processing of
mammograms [7]. Hence, identification and segmentation
of the pectoral muscle is essential.
190
J Med Syst (2017) 41:190
Page 2 of 13
Fig. 1 Estimated cases (dark
black bar) and deaths (light bar)
of breast cancer yearwise, a for
both sex male and female, b for
female, c for male
(a)
(b)
(c)
In the Computer-Aided Detection (CAD), there are
three anatomical landmarks i.e. pectoral muscle, breast
boundaries and nipple [8]. The current work is mainly
concentrated on emphasizing the accuracy of pectoral muscle identification and segmentation. Since, mammographic
parenchyma and pectoral muscle has similar characteristics,
which can be a source for misdiagnosis and cause of high
false positve rate (FP) for breast cancer detection. There
is large variation in texture, size, position, intensity and
shape of the pectoral muscle, due to patient positioning during mammograms acquisition [8]. Therefore, detection of
pectoral muscle is important and challenging task.
To overcome the problems and improve the segmentation
of pectoral muscle, various approaches are proposed and can
be found in the literature [9–18]. Chakraborty et al. [9] put
forward a weight function approach to enhance the pectoral
boundaries and then local gradient to detect the pectoral
edge points. But, False Positive (FP) rate with [9] is greater
than 4.22%. This method is not suitable for pectoral covered with similar tissue at upper part of the pectoral muscle.
The new method was introduced by Li Liu et al. [10] using
local non-Gaussianity measure, which used pixel wise statistical features for detection of pectoral muscle. Sreedevi
S et al.[11] proposed a method for noise removal and pectoral muscle segmentation of mammograms, using Discrete
Cosine Transform (DCT) based non local mean filter and
canny edge detection in combination with global threshold
respectively. To suppress the pixel connected outside breast
region, connected component labeling approach was used in
their work. The method was tested for 161 mammograms
from MIAS database, with accuracy of 90.06%. This technique have limited effect for pectoral having more than one
layer. The novel method was developed by R. J. Ferrari
et al.[12] for automatic pectoral muscle segmentation, using
multiresolution approach based on Gabor wavelet. The algorithm was tested and evaluated on 84 mammograms for edge
detection of pectoral muscle from MIAS database. Where,
performance evaluation of the method [12], was computed
in terms of false positive (FP) 0.58% and false negative
(FN) 5.77% respectively, with mean Hausdorff distances of
3.84 mm. In this method segmentation of pectoral muscle is
difficult, when pectoral muscle is hidden in the dense tissue.
The combination of bit depth reduction and wavelet analysis approach for pectoral muscle detection was proposed
by Mario Mustra et al. [13]. The proposed method [13] was
tested for 40 mammograms with an accuracy of 85%. This
method is suitable only, when the contrast of pectoral muscle is higher than that of surrounding tissue. Yanfeng Li
et al. [14] put forward a approach based on homogeneous
texture and high intensity deviation, features of pectoral
region and kalman filter to identify and smoothing the edges
of pectoral muscle respectively. The approach was tested for
322 and 100 mammographic images from MIAS and Digital
Database for Screening Mammography (DDSM) database
with acceptance rate of 90.06% and 92% respectively. This
approach have limited effect on low contrast mammograms.
Wang et al. [15] presented a method for edge detection of
pectoral muscle based on a active contour model and an
discrete time Markov chain (DTMC). This method can overcome the disadvantages such as double layer muscles and
J Med Syst (2017) 41:190
line edges. But, the method is not suitable for segmentation
of small pectoral muscle. The pectoral muscle segmentation
based on Minimum Spanning Trees (MST) and Adaptive
Pyramids (AP) was introduced by F. Ma et al. [16]. However, the results with this method are not satisfactory for
multi-layer pectoral muscle. David Raba et al. [18] suggested an algorithm to segment pectoral muscle, based on
combination of region growing and morphological operation. If the contrast across breast tissue and pectoral muscle
are indistinct, their may be over-segmentation of pectoral
muscle. Sze Man Kwok et al.[17] proposed a method to
detect the pectoral muscle based upon Hough transform and
cliff detection. In this method, straight line concept was
used first, to detect the edge of pectoral muscle roughly,
and then polished to a curve. But due to complicated texture present in pectoral muscle, it is difficult to find correct
muscle boundaries in some cases.
Considering above facts, in this study, we have developed
a simple intensity based detection algorithm for automatic
pectoral boundary detection and pectoral muscle segmentation. The algorithm starts by convolution of mammograms
with proposed intensity based filter enhancing the Region of
Interest (ROI) having pectoral muscle. Finally, curve fitting
technique is used to refine the boundary of pectoral muscle
based on the boundary points obtained after enhancement
processes using threshold technique.
The paper is organized as follow. “Introduction”, presents
introduction with current literature studies. “Methodology”
describes the concept of proposed approach for pectoral
muscle detection and segmentation from mammograms.
The quantitative performance evaluation is explained in the
third section. Experimental results and discussion based
on the proposed method are demonstrated in the fourth
and fifth section. Conclusion drawn based on the above
experimentation is given in last section.
Methodology
The distinct variation in intensity between pectoral region
and breast tissue, with approximately triangular structure,
and gently taper from top to bottom are some of the noteworthy anatomical features of the pectoral muscle [14, 19,
20]. An enhancement filter and frontier evolution approach
based on the above listed features is proposed in this paper
to detect the boundary and segment pectoral muscle automatically. The radiopaque artifacts i.e. (labels, wedges, etc.)
as present in Fig. 2a, can affect the detection of pectoral
muscle, hence removal of artifact from the mammogram is
important and achieved using pre-processing. To locate the
pectoral muscle on the top left corner of the image, all left
oriented mammograms were flipped based on mean computation. To achieve this vertical centroid and mean of the
Page 3 of 13 190
(a)
(b)
Fig. 2 a Original mammogram with artifacts, case mdb051, b preprocessed mammogram
mammograms were computed of the left and right half. If
the left half mean is greater than right half mean, flip the
mammogram else retain [21].
Pre-processing
The pre-processing is followed by artifact suppression i.e.
(labels, wedges) from mammographic images using following steps:
1. To reduce the noise present in the mammographic
images, median filtering approach is used, as this technique preserve the contrast and does not shift the
boundaries of the image.
2. Convert the filtered mammogram to binary using
threshold technique, providing number of objects (
artifacts and breast region) as an output.
3. The breast region is obtained by computing the size of
each object from 2, considering only large object and
removing small objects from the image.
4. As output of 3 is in binary form and contains only breast
and pectoral region have pixel value 1(white). Whereas,
background contains pixel value 0 (black).
5. The pixel value 1 (white) is replaced by corresponding
pixel value of original image in the binary mask generated in 4, to extract the original pixel values of the
mammograms, contain pectoral and breast region only
as seen in Fig. 2b denoted as, Ib .
6. The output Ib , of 5 is used to identify and extract
pectoral muscle using proposed approach.
Intensity based enhancement filter
In the mammograms, pectoral muscle have high gray level
intensities and approximately triangular shape. Based on
the previous comprehension, an enhancement mask is frequently used to enhance the pectoral muscle during processing for mammograms. In this work, to increase the intensity
190
J Med Syst (2017) 41:190
Page 4 of 13
less than pixel value present in the pectoral muscle, while
processing mammograms.
(a)
Pectoral muscle boundary point selection
(b)
Fig. 3 a Linear enhancement filter, b Linear enhancement filter
rotated by 90◦ clockwise
of gray level of the pectoral muscle a linear enhancement
mask with few coefficients is proposed and given as Eq. 1.
Ie = y(i, j )
= Cp,q
N M
(Ib (i + p, j − q) − Ib (i + p, j + q))
p=0 q=1
+Cc Ib (i, j ).
(1)
where, Ie is the enhanced pectoral mammographic image,
Ib (i, j ) is the pixel intensity of an image at point (i, j ), N
are the number of image columns. The weighted differentiation effect for number of pixel pairs along vertical direction
is denoted as M, Cp,q and Cc are filter coefficients. The
mammogram is convoluted with linear enhancement filter
as shown in the Fig. 3, to implement the above equation. As
pectoral muscle is gently tapers from top to bottom, the filter shown in Fig. 3a is rotated by 90◦ clockwise as shown
in the Fig. 3b, so as to emphasize high intensity pixels
from mammograms of pectoral region as shown in Fig. 4b.
As, pectoral muscle have higher pixel intensity values compared to breast tissue with approximately triangular shape,
which can be clearly depicted from Fig. 4a. The parameters Cp, q of the enhancement filter were set to 1 and -1
(except center coefficient Cc ) to have the zero sum of coefficients. The parameter Cc provides the intensity contribution
of the current pixel’s. Usually, higher the value of Cc leads
to superior enhancement of the pectoral region as depicted
in Fig. 4b. The center point coefficient Cc is chosen in the
range, 0 < Cc < 2, so as to retain the pixel values that are
Based on the intensity difference characteristic of pectoral
region and breast tissue, Jawad Nagi et al. [22] found four
points iteratively on the pectoral boundaries using seed
region growing approach and straight line equation for segmentation of the pectoral muscle. However, it is not easy
to segment the pectoral muscles accurately, having variation in shapes using straight line concept. In order to obtain
the accurate boundaries, intensity difference characteristic
of pectoral and breast region is used for selecting the boundary points of pectoral muscle on the mammograms. Then
curve fitting technique is used to segment the accurate shape
of pectoral muscle. The pectoral boundary point search and
selection is explain in steps below:
1. Find the maximum gray value of the enhanced image Ie
and use the same as threshold value T .
2. Define M rows and N columns for the enhanced
mammogram, scan row-wise with specific pixel interval from top to bottom, so as to obtain the pectoral
boundary points.
3. Select and scan entire columns of first row from M
rows, and store the first pixel location value of the row
and column from (M, N) having intensity value less
than threshold value T .
4. Increment the value of row with specific pixel interval
toward downside of the mammogram, and repeat step.
3 for entire image.
5. The output of the step. 4 provides the location points
on the pectoral boundaries, as depicted in Fig. 4c on
pre-processed original mammogram with dots.
6. The location points obtained in step. 4 and shown
in step. 5 are connected to detect the rough edge of
pectoral muscle from mammographic images.
Fig. 4 a Pre-process original
mammogram case mdb001, b
enhanced mammographic image
using linear enhancement filter,
c pectoral boundary points
obtained based on intensity
variation using threshold (dots)
(a)
(b)
(c)
J Med Syst (2017) 41:190
Page 5 of 13 190
Taking derivative of error Er with respect every coefficient P0 , Pk for k = 1, 2, 3. and set each equals to zero
as below to minimize Eq. 4.
Least squares error approach
The location data points obtained on pre-process original mammogram (xi , yi ) Fig. 4c dots, with above process
can vary in some cases, due to intensity variation in pectoral region, and may lead to squiggly boundaries, during
pectoral muscle segmentation.
Hence, to overcome this problem curve fitting is used
in this work to refine the boundaries of the pectoral muscle [23]. The polynomial curve fitting with least squares
error method as shown with an example in Figs. 5 and 6
for case mdb025, is introduced to segment accurate shape
of the pectoral muscle [24]. Figures 5a, b and 6b shows the
actual data points (circles) obtained using boundary point
selection. Whereas, Figs. 5a, b and 6c depicts detected pectoral boundary (dotted line) of pectoral muscle connecting
actual data points. But, this may leads in irregular pectoral
boundary detection, causing increase in False Positive Rate
(FPR). Hence, to overcome this drawback curve fitting LSE
method is used and is presented in Figs. 5a, b and 6d, e
with continuous line for accurate boundary detection of pectoral muscle. However, the concept of LSE is highlighted
in Figs. 5b and 6d showing difference with (white dotted
line in Fig. 6d) between actual data point (dotted line) and
obtained using curve fitting (continuous line) , representing actual pectoral muscle segmentation, as seen in Fig. 6f.
However, 3rd order polynomial is used for detection of pectoral boundaries, as it detects varying shape accurately as
give below:
The polynomial equation of 3rd order is give as Eq. 2.
F (x) = P0 +
3
Pk x k
(2)
k=1
where, P0 and Pk are the unknown coefficients for k =
1, 2, 3. The ‘best’ curve fit can be obtained, when error
between the fitted line and data points is minimum.
The least squares error approach is used for error computation, and its general form is given as per Eq. 3.
n
(Di )2 =
(yi − F (xi ))2
Er =
(3)
i=1
where Er is the error, n are the number of data points,
F (xi ) are the fit functions at each data point given by
Eq. 2, and yi is data, substituting Eq. 2 in Eq. 3 to obtain
Eq. 4 i. e. error equation for 3rd order.
Er =
n
i=1
yi − P0 +
3
2
Pk x k
(4)
k=1
where, ‘3’ is the order of polynomial and i the sum of
current data point.
n
3
∂Er
yi − P0 +
=0
= −2
Pk x k
∂P0
i=1
k=1
n
3
∂Er
= −2
Pk x k
yi − P0 +
x=0
∂P1
i=1
k=1
n
3
∂Er
k
= −2
Pk x
yi − P0 +
x2 = 0
∂P2
i=1
k=1
n
3
∂Er
k
yi − P0 +
x 3 = 0 (5)
= −2
Pk x
∂P3
i=1
k=1
The above Eq. 5 can be presented in matrix form as
Eq. 6 to obtain the values of unknown coefficients.
2 3 ⎤⎡ ⎤ ⎡ ⎤
n
x
x
x
P0
yi
2i i3 i4
⎢ xi
⎢ ⎥ ⎢
⎥
xi
xi
xi ⎥
⎢
⎥ ⎢ P1 ⎥ ⎢ (xi , yi ) ⎥
⎣ x 2 x 3 x 4 x 5 ⎦ ⎣ P2 ⎦ = ⎣ (x 2 , yi ) ⎦ (6)
i3
i3 i4 i5 i6
P3
(xi , yi )
xi
xi
xi
xi
⎡
A
X
B
where, (xi , yi ) are data points, for i = 1, 2, ..., n and all
above summations are over i, P0 , Pk are the unknown
coefficients to be obtained, for k = 1, 2, 3.
The Eq. 6 is now in linear form with respect to coefficients. Thus, the unknown coefficients P0 , Pk can be
found using following Eq. 7.
AX = B
Where, X is the coefficient matrix solved as below.
X = A−1 ∗ B
(7)
Pectoral muscle segmentation
The segmentation of the pectoral muscle is obtained based
on the ROI, separated in two parts pectoral region (ROI)
and breast region with black line as seen in Fig 6e using
function roipoly, used to process ROI, without affecting
the information of breast region [25]. The new data points
(xi , yi ) obtained after applying curve fitting approach on
the mammogram (Ib ), are taken into account for selecting
the candidates of pectoral region using Eq. 8.
Ipd (x, y) = roipoly(Ib , xi , yi )
(8)
where, Ipd (x, y) denotes the binary mask of image size,
having pixel value ‘1’ for pectoral region or ROI and ‘0’
otherwise, as shown in Fig. 7a. Furthermore, pectoral region
190
J Med Syst (2017) 41:190
Page 6 of 13
(a)
(b)
Fig. 5 Least squares error method (LSEM) for best fit, a Actual data points circle (x, y) and curve fitting data point brown (x, F (x)), b difference
between actual and curve fitting data points using LSEM, e.g. case mdb025
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 6 a Pre-processed original mammogram case mdb025, b data
points (circle) obtained using enhancement filter, c pectoral boundary
detected using data points (dotted line), d pectoral boundary detected
using data points (dotted line) and curve fitting (continuous line)
highlighting difference with (dotted white line) between data points
and curve fitting, e pectoral boundary detected using curve fitting (continuous line), f pectoral muscle segmented based on curve fitting using
LSEM
J Med Syst (2017) 41:190
Page 7 of 13 190
(a)
(b)
Fig. 7 a Generated binary mask using function roipoly, case mdb025
b pectoral muscle segmentation after applying curve fitting technique
segmented mammogram as depicted in Fig. 7b, is achieved
using Eq. 9.
Ips (i, j ) =
0,
Ib (i, j ),
Ipd (i, j ) == 1
elsewhere.
(9)
where, Ips denotes the pectoral muscle segmented mammogram, that can be used for post processing of mammograms
for mass detection.
Quantitative performance evaluation
To evaluate the performance author has drawn the groundtruth for pectoral muscle manually, and separately verified
by two radiologist experts, for each MLO mammograms for
quality evaluation of the pectoral muscle detection. Three
parameters, Hausdorff distance Hd , False Positive (FP) rate
and False Negative (FN) rate are used to evaluate the accuracy quantitatively of the proposed approach for automatic
pectoral muscle detection [8].
Fig. 8 Schematic representation of automatic pectoral detection
system
Table 1 Details of the database
Database
Mini-MIAS
Projections
Spatial resolution
Gray-level quantization
Dimension
Digitizer
Database Size
Database Tested
MLO
50 μm/pixel
8 bits
1024×1024 pixels
Joyce-Loebl microdensitometer SCANDIG-3
322
320
Hausdorff Distance (Hd ): The similarity between
detected and ground-truth set is determine using Hausdorff distance, and is given by Eq. 10.
Hd (A1 , A2 ) = max{h(A1 , A2 ), h(A2 , A1 )}
(10)
where h(A1 , A2 ) = max min a1 − a2 where, A1 ∈
a1 ∈A1 a2 ∈A2
set of detected pectoral region, A2 ∈ set of groundtruth pectoral region and . is the Euclidean distance
between a1 and a2 .
False positive (FP) pixels: The pixels detected in pectoral muscle other than or outside ground-truth or
190
J Med Syst (2017) 41:190
Page 8 of 13
Fig. 9 Small size pectoral
muscles detected (black line)
using proposed method for
different cases of MIAS
database a mdb050 b mdb106 c
mdb109 d mdb148 e mdb148 f
mdb299
(a)
(b)
(d)
(e)
reference region, and is computed using Eq. 11 in
percentage as below.
Percentage FP pixel =
|A1 ∪ A2 | − |A2 |
× 100%(11)
|A2 |
False negative (FN) pixels: The pixels outside the
detected pectoral region but inside the ground truth
(c)
(f)
region, and is computed using Eq. 12 in percentage as
below.
Percentage FN pixel =
|A1 ∪ A2 | − |A1 |
× 100%(12)
|A2 |
where, A1 ∈ detected pectoral region pixels set and
A2 ∈ reference or ground-truth pectoral region pixels
set.
Fig. 10 Pectoral muscles
detected with varying shape
(black line) using proposed
method for different cases of
MIAS database a mdb004 b
mdb111 c mdb172 d mdb260 e
mdb178
(a)
(b)
(c)
(d)
(e)
J Med Syst (2017) 41:190
Page 9 of 13 190
Fig. 11 Pectoral muscles
detected having low contrast
(white line) using proposed
method for different cases of
MIAS database a mdb065 b
mdb066 c mdb138 d mdb284
(a)
Experimental results
The experimental results obtained using proposed algorithm
for segmentation of pectoral muscle is described in this
section. The proposed algorithm were tested using digitized mammograms with 200 μm spatial resolution form
mini-MIAS dataset. The details of the database is presented
in the Table 1 (http://marathon.csee.usf.edu/Mammography/
Database.html). To locate the pectoral on the left top corner of mammogram all left oriented mammograms were
flipped. We have processed 320 mammograms using linear
enhancement filter, since pectoral muscle was not detected
for two mammograms by radiologist. The proposed method
elaborated in Fig. 8, has correctly detected and segmented
pectoral muscle from mammograms verified by radiologists, illustrated with some examples.
Figures 9, 10, 11, 13 and 14 demonstrate the results
for some pectoral muscle detection of MLO mammograms
having small size, varying shapes, low contrast, multiple
or double layer and dense mammograms using proposed
(b)
(c)
(d)
technique. To assess the quality of muscle detection algorithm, the edges were manually drawn on each mammogram
by the author and verified by two radiologist separately.
These manually drawn muscle boundaries on the MLO view
of mammogram were used as reference ground truth. The
consensus was reached for the changes exist after discussion
with radiologists. As is apparent from Fig. 9a–f, proposed
method detects (black line) the small pectoral region accurately for different mammograms. Figure 10a–e present the
results for different shapes and texture (black line) using
proposed approach. The low contrast muscle detection is
evident in Fig. 11 (white line), the proposed method gives
good results for low contrast pectoral mammograms, having distinct intensity feature between pectoral region and
breast tissue, but may lead in either false negative or positive
results for the mammograms with same intensity or overlap
of breast tissue and muscle i.e (glandular mammograms).
Which is elaborated with an e.g. case mdb065 demonstrated
in Fig. 12a–d step wise, where, Fig. 12 (a) is the original mammogram, (b) shows the result obtained for muscle
Fig. 12 a Original mammogram
case mdb065 b pectoral muscle
segmented by proposed method
c pectoral muscle segmentation
as per ground truth d similarity
measure of ground truth and
proposed method pectoral
muscle indicating false positive
region
(a)
(b)
(c)
(d)
190
J Med Syst (2017) 41:190
Page 10 of 13
Fig. 13 Pectoral muscles
detected (black line) having
double layer using proposed
method for different cases of
MIAS database a mdb034 b
mdb110 c mdb123 d mdb185
(a)
(b)
segmentation using proposed method, (c) presents the result
of ground-truth pectoral region and (d) depicts the similarity match of the result obtained using proposed method
and reference pectoral muscle. The proposed method effectively enhance and detect the low contrast muscle using
linear enhancement filter but leads to have false negative
region across the overlap of muscle and breast tissue having
fibro-glandular region. The proposed method gives better results, that can be seen clearly from Fig. 13, (black
line) for different mammograms having multiple or double layer muscles. Figure 14 demonstrate the ability of the
proposed method to detect pectoral muscles from dense
mammograms with varying muscle shapes. The results
(c)
(d)
of pectoral muscle detection were categorized into three
classes:
•
•
•
Accurate: In case of accurate results, pectoral muscle
detected was similar to that of ground-truth.
Acceptable: In case of acceptable results, more than
60% of pectoral region detected was from the groundtruth, with limited difference in case of glandular tissue
present at the lower part of the mammogram.
Inaccurate: The rest of the results were inaccurate and
not correctly detected.
The accurate or acceptable rate achieved for pectoral muscle
detection is 96.56% with proposed approach, tested on 320
Fig. 14 Pectoral muscles with
different shape, size, contrast
detected (black line) using
proposed method for different
cases of MIAS database a
mdb151 b mdb190 c mdb215 d
mdb240 e mdb224
(a)
(b)
(c)
(d)
(e)
J Med Syst (2017) 41:190
Page 11 of 13 190
Table 2 Boundary detection result classification for MIAS database
Classified
category
Number of
mammograms
Results in
Percentage(%)
Accurate
Acceptable
Inaccurate
279
30
11
87.19
9.37
3.44
mammograms from MIAS database, presented in Table 2.
Moreover, ground truth of 84 mammograms were used from
MIAS database, which was provided by R. M. Rangayyan
[12] were selected for quantitative performance evaluation,
in terms of mean Hd , FPm and FNm rate were 3.79 mm,
0.93% and 5.07% respectively.
Discussion
The proposed technique requires the information and prior
knowledge about the intensity and shape feature of the pectoral region for detection of the pectoral muscle. The pectoral region appears with high gray level intensity roughly
occupies the triangular area on the top corner of the mammogram. The approach based on intensity characteristic and
threshold technique for boundary point selection has been
used for detection of pectoral muscle.
To enhance the pectoral region a filter mask was designed
with few coefficients. The designed filter mask considers
the transition intensity variation, across the pectoral muscle. This is the main significance of the designed filter
mask over the conventional filter mask. The intensity difference between the breast and pectoral region of enhanced
mammogram has been used to identify the correct pectoral
boundary points using threshold technique. Further more all
the boundary points obtained as shown in Fig. 6b (dots)
are connected to obtain the pectoral boundary line. The
center Cc value of pixel kernel of 3× 2 is consider high,
since high Cc value can effectively enhance the low contrast region, useful in detection of low contrast and multi or
double layer pectoral muscle, as shown in Figs. 11 and 13a–
d. The location and high gray level intensity of the pectoral
muscle’s has been used as a prior knowledge for detection.
The examples presented in paper for small size, varying
shape, low contrast and double layer shows the effectiveness
and advantage of the proposed approach. In some cases,
accurate pectoral muscle detection is difficult, due to similar gray level intensity of pectoral and breast region, is the
only limitation of proposed technique for e.g. Fig. 12. The
comparison for pectoral muscle detection using proposed
technique and method introduced by Vikhe and Thool [26]
is demonstrated in Fig. 15. Figure 15a depicts the results
obtained using method [26] for two cases with case number (i.e. mdb028 and mdb156). Whereas, Fig. 15b presents
the results obtained using proposed technique for the same
cases. The result obtained using method [26] shows irregular pectoral boundary detection. However, Fig. 15b depicts
the correct detection of pectoral with proposed technique.
Furthermore, example of pectoral muscle detection using
proposed and Chakraborty’s [9] method is presented in
Fig. 16. The results obtained using Chakraborty’s method
for case mdb097 and mdb156 is demonstrated in Fig. 16b,
d. The obtained result shows that method [9] fails for small
pectoral and pectoral covered with similar tissue at upper
part of the pectoral muscle, leading in over segmentation
of pectoral region. Whereas, Fig. 16 a, c depicts the correct
detection of pectoral using proposed approach.
The mean Hausdorff Distance (Hd in mm), FPm and
FNm rate, parameters are computed for performance analysis to show the effectiveness of the proposed intensity based
approach. As mean FP and FN rate are the key parameters
for quantitative analysis of pectoral muscle segmentation.
The FPm rate of 3.71 and FNm rate of 5.95 were achieved
with Adaptive pyramid method presented by F. Ma et al.
[16]. The shape based enhancement mask method proposed
by C. Chen et al. [8] have obtain FPm and FNm rate of
1.02 and 5.63 respectively with Hd of 3.53±1.61 tested
on MIAS database. However, performance obtained using
proposed approach in terms of mean FPm and FNm rate
is i.e. 0.93 and 5.07 respectively, with Hd of 3.79±1.67,
which is better compare to mean FPm and FNm rate (i.e 1.56
Fig. 15 Pectoral muscles
detected (black line) using a
method proposed by Vikhe and
Thool [26] for case mdb028 and
mdb156 from MIAS database b
proposed approach for case
mdb028 and mdb156 from
MIAS database
(a)
(b)
(a)
(b)
190
J Med Syst (2017) 41:190
Page 12 of 13
Fig. 16 Pectoral muscles
detected (black line) using
proposed method for case of
MIAS database a mdb097 and d
mdb156 and by using
Chakraborty’s [9] method for
case b mdb097 and c mdb156
(a)
(b)
(c)
(d)
and 6.93) achieved using method [26] and methods [8, 16].
Hence, as per the opinion of radiologists, results obtained
using proposed technique for pectoral muscle detection are
better and can be helpful for pre-processing in computerized
analysis.
Compliance with Ethical Standards
Conclusion
References
This paper deals with the boundary detection and segmentation of pectoral muscle from the mammograms. The
proposed approach consists of pectoral region enhancement
based on intensity based feature using linear filter mask
for detection of the pectoral boundaries. The method presented in this paper does not depend on detection of straight
line. The intensity based designed mask effectively enhance
pectoral region based on the characteristics of pectoral
muscle first. Then, pectoral boundary point is determined
based on intensity difference between enhanced pectoral
and breast region receptively using threshold technique, to
obtain the rough pectoral boundaries. Finally, refining of
the rough pectoral boundaries using LSE method is used
for accurate pectoral muscle segmentation. The experimental results obtained for pectoral muscle detection prove that
the proposed method gives good performance, with 96.56%
accuracy rate, tested on 320 mammograms form mini-MIAS
database, as per the opinion of expert radiologists.
Future research work will concerns with testing of mammograms from Digital Database for Screening Mammography (DDSM) database and overcome the limitation of
pectoral muscle detection, due to similar intensity of pectoral and breast region of the mammogram to increase the
accuracy of intensity based method.
1. Tang, J., Rangayyan, R. M., Xu, J., El Naqa, I., and Yang,
Y., Computer-aided detection and diagnosis of breast cancer
with mammography: recent advances. IEEE Trans. Inf. Technol.
Biomed. 13(2):236–251, 2009.
2. Vikhe, P. S., and Thool, V. R., Contrast enhancement in mammograms using homomorphic filter technique. In: International conference on signal and information processing (IConSIP), pp. 1–5:
IEEE, 2016.
3. American Cancer Society, Cancer facts and figures, American
Cancer Society, Atlanta, GA, 2005–2015.
4. Ganesan, K., Acharya, U., Chua, C. K., Min, L. C., Abraham,
K., and Ng, K., Computer-aided breast cancer detection using
mammograms: a review. IEEE Rev. Biomed. Eng. 6:77–98, 2013.
5. Agrawal, P., Vatsa, M., and Singh, R., Saliency based mass detection from screening mammograms. Signal Process. 99:29–47,
2014.
6. Kalra, P. K., Kumar, N., et al., An automatic method to enhance
microcalcifications using normalized tsallis entropy. Signal Process. 90(3):952–958, 2010.
7. Ganesan, K., Acharya, U. R., Chua, K. C., Min, L. C., and
Abraham, K. T., Pectoral muscle segmentation: a review. Comput.
Methods Prog. Biomed. 110(1):48–57, 2013.
8. Chen, C., Liu, G., Wang, J., and Sudlow, G., Shape-based automatic detection of pectoral muscle boundary in mammograms. J.
Med. Biol. Eng. 35(3):315–322, 2015.
9. Chakraborty, J., Mukhopadhyay, S., Singla, V., Khandelwal, N.,
and Bhattacharyya, P., Automatic detection of pectoral muscle
using average gradient and shape based feature. J. Digit. Imaging
25(3):387–399, 2012.
10. Liu, L., Liu, Q., and Lu, W., Pectoral muscle detection in
mammograms using local statistical features. J. Digit. Imaging
27(5):633–641, 2014.
11. Sreedevi, S., and Sherly, E., A novel approach for removal of pectoral muscles in digital mammogram. Procedia Computer Science
46:1724–1731, 2015.
12. Ferrari, R. J., Rangayyan, R. M., Desautels, J. L., Borges, R., and
Frere, A. F., Automatic identification of the pectoral muscle in
mammograms. IEEE Trans. Med. Imaging 23(2):232–245, 2004.
Acknowledgments The authors would like to thank Dr. Sushil
Kachewar, Associate Professor, Dr. Sham, Assistant Professor, Dr.
Rahul Umbarkar and their team, Department of Radiodiagnosis and
Imaging of Rural Medical College, Pravara Institute of Medical Science (PIMS), Loni (Deemed University), for providing their timely
consultation.
Ethical approval This article does not contain any studies with
human participants or animals performed by any of the authors. The
authors declare that they have no conflict of interest.
J Med Syst (2017) 41:190
13. Mustra, M., Bozek, J., and Grgic, M., Breast border extraction
and pectoral muscle detection using wavelet decomposition. In:
EUROCON 2009, EUROCON’09. IEEE, pp. 1426–1433: IEEE,
2009.
14. Li, Y., Chen, H., Yang, Y., and Yang, N., Pectoral muscle segmentation in mammograms based on homogenous texture and intensity deviation. Pattern Recognition 46(3):681–691,
2013.
15. Wang, L., Zhu, M.-l., Deng, L.-p., and Yuan, X., Automatic
pectoral muscle boundary detection in mammograms based on
markov chain and active contour model. J. Zhejiang Univ. Sci. C
11(2):111–118, 2010.
16. Ma, F., Bajger, M., Slavotinek, J. P., and Bottema, M. J.,
Two graph theory based methods for identifying the pectoral
muscle in mammograms. Pattern Recogn. 40(9):2592–2602,
2007.
17. Kwok, S. M., Chandrasekhar, R., Attikiouzel, Y., and Rickard,
M. T., Automatic pectoral muscle segmentation on mediolateral oblique view mammograms. IEEE Trans. Med. Imaging
23(9):1129–1140, 2004.
18. Raba, D., Oliver, A., Martı́, J., Peracaula, M., and Espunya, J.,
Breast segmentation with pectoral muscle suppression on digital mammograms. In: Pattern recognition and image analysis,
pp. 471–478: Springer, 2005.
Page 13 of 13 190
19. Mustra, M., and Grgic, M., Robust automatic breast and pectoral muscle segmentation from scanned mammograms. Signal
Process. 93(10):2817–2827, 2013.
20. Bora, V. B., Kothari, A. G., and Keskar, A. G., Robust automatic
pectoral muscle segmentation from mammograms using texture
gradient and euclidean distance regression. J. Digit. Imaging
29(1):115–125, 2016.
21. Vikhe, P. S., and Thool, V. R., Mass detection in mammographic images using wavelet processing and adaptive threshold
technique. J. Med. Syst. 40(4):1–16, 2016.
22. Nagi, J., Abdul Kareem, S., Nagi, F., and Ahmed, S. K., Automated breast profile segmentation for roi detection using digital
mammograms. In: IEEE EMBS Conference on Biomedical Engineering and Sciences (IECBES), 2010, pp. 87–92: IEEE, 2010.
23. Mathews, J. H., and Fink, K. D., Numerical methods using MATLAB. Vol. 31. Upper Saddle River, NJ: Prentice Hall, 1999.
24. Sarfraz, M., and Khan, M., An automatic algorithm for approximating boundary of bitmap characters. Futur. Gener. Comput.
Syst. 20(8):1327–1336, 2004.
25. McAndrew, A., A computational introduction to digital image
processing, 2nd edn. CRC Press, 2015.
26. Vikhe, P. S., and Thool, V. R., Intensity based automatic boundary identification of pectoral muscle in mammograms. Procedia
Computer Science 79:262–269, 2016.
Документ
Категория
Без категории
Просмотров
0
Размер файла
3 146 Кб
Теги
017, 0839, s10916
1/--страниц
Пожаловаться на содержимое документа