A Fuzzy Inference System Combined with Wavelet

Transform for Breast Mass Classification

Pelin Görgel, Ahmet Sertbas, and Osman N. Ucan

 This paper is organized as follows: In Section II, Fast

Abstract—This paper proposes a combination of the Fast
Wavelet Transform (FWT) and Adaptive Neuro-fuzzy Inference
System (ANFIS) methods. The goal is classification of breast
masses as benign or malignant by applying this method
consecutively to the extracted features of the Region of Interests
(ROIs). This study is developed to decrease the number of the
missing cancerous regions or unnecessary biopsies. The neuro-
fuzzy subtractive clustering classification method achieved a
classification accuracy of 85% without using FWT multi-
resolution analysis and 92% with FWT. The satisfying results
demonstrate that the developed system could help the
radiologists for a true diagnosis.
Keywords—Breast cancer, mass classification, ANFIS, fast
wavelet transform.
I. INTRODUCTION
B REAST cancer is one of the most life-threatening
diseases especially in developed countries. If it is
detected at an early stage, the chance of survival rises.
Although mammography provides an advance imaging for
breast cancer, the unnoticeable masses could sometimes be
missed. This causes mortalities due to the late detection and
treatment [1]. The Computer Aided Diagnosis (CAD)
systems using digital mammography have recently shown a
great potential in providing a second eye to the radiologists
for a correct malignant-benign decision. A CAD system was
developed by Delogu et al. [2] for the classification of
mammographic masses as malignant or benign. They used
twelve features based on shape, intensity and size of the
segmented masses. In the study by Rangayyan et al. [3]
combined speculation index, three shape factors, fractional
concavity and compactness and achieved classification
accuracy of 81.5%. Cascio et al. [4] used geometrical features
about shape parameters for each region of interest to classify
the masses. They used supervised neural network which
achieved a sensitivity value of 82%. Tralic et al. [5]
calculated three shape factors, namely Fourier descriptors,
compactness and moments. Classification was performed
using both single layer and multilayer perceptron neural
networks and the highest accuracy was 91.5%.
Manuscript received February 11, 2012.
P. Görgel is with the Computer Engineering Department, University of
Istanbul, Istanbul, Turkey (corresponding author to provide phone: 212-473-
7070; fax: 212-473-7044; e-mail: paras@istanbul.edu.tr).
A. Sertbas is with the Computer Engineering Department, University of
Istanbul, Istanbul, Turkey. (e-mail: asertbas@istanbul.edu.tr).
O. N. Ucan is with the Electrical Engineering Department, University of
Aydin, Istanbul, Turkey. (e-mail: uosman@aydin.edu.tr).
Wavelet Transform which is used before the feature
extraction is explained. Besides the subtractive clustering
expressions and ANFIS architecture are stated. In Section III
the used mammogram masses data set is mentioned and the
obtained experimental results are presented and discussed.
Finally Section IV draws the conclusion and gives some final
remarks.
II. MATERIALS AND METHODS
A. Fast Wavelet Transform (FWT)
Wavelets are counted as a powerful signal processing
foundation of Mallat [6] in 1987. The Fast Wavelet
Transform is a computationally efficient form of the discrete
wavelet transform (DWT) [7]. It is a multi-resolution analysis
method that provides frequency decomposition of the images
or signals using scaling ( M j ,k ( x) ) and wavelet (\ j , k ( x ) )
functions.
M j ,k ( x ) 2 j / 2 M (2 j x  k ) (1)
\ j ,k ( x ) 2 j / 2\ ( 2 j x  k ) (2)
In the above equations j and k determines the scaling
and wavelet functions’ width and the position respectively
while the value 2 j / 2 controls the amplitude. (3) and (4)
illustrate the approximation and detail coefficients
respectively in the two-dimensional wavelet transform.
f ( x, y ) is used for the image and m and n are for the image
size. The index i is H for horizontal, V for vertical and D
for diagonal details.
M 1N 1
¦ ¦ f ( x , y ) M j , m , n ( x, y )
1
WM ( j , m, n) (3)
MN x 0y 0
M 1N 1
¦ ¦ f ( x, y)\ i
1
W\i ( j , m, n) ( x, y ) (4)
MN j ,m,n
x 0y 0
The FWT is implemented via digital filters and
downsamplers as formulated in (5) and (6). After a FWT,
four sub-images one of which is the approximation image and
the others are horizontal, vertical and diagonal detail images
are obtained. Low pass ( hM (n) ) and high pass ( h\ (n) ) filters
are used for the approximation and detail coefficients
respectively. After the filtering step, downsampling is
implemented for the scale changing as seen in Fig. 1. The fast

978-1-4673-1118-2/12/$31.00 ©2012 IEEE 644 TSP 2012

 boundary based Mean Center-Border Distance representing
h\ (n ) W\ ( j  1, n ) the similarity between a circle and the ROI and the other is
f (n) shape based Symmetry. These calculated numeric features
WM ( j  2, n )
listed below provide feature matrices to the fuzzy inference
h\ (n ) W\ ( j, n ) system.
C. Fuzzy subtractive clustering
hM (n ) WM ( j  1, n )
The subtractive clustering is used to determine the number
of clusters of the data being proposed, and then generates a
hM (n ) WM ( j, n )
fuzzy model [8]. The purpose of this algorithm is to estimate
both the number and initial locations of cluster centers [9].
Fig. 1. A two scale Fast Wavelet Transform
The subtractive clustering method partitioned the training
TABLE 1.
data into groups called clusters. By the end of clustering, a set
Extracted features of fuzzy rules will be obtained. The FIS is generated with
minimum number of rules. The clustering is carried out in a
Area Bounding Box multidimensional space; the related fuzzy sets must be
Centroid Filled Area obtained. Let the cluster set be Z1 , Z 2 ,...Z n for n data. The
Extrema Convex Hull
Eccentricity Major Axis Length subtractive clustering algorithm steps are as follows:
Orientation Minor Axis Length 1.The initial potential value for each data point
Symmetry Mean Center-Border Distance
Solidity Equiv Diameter ( Z i ) as in (7) is computed.
Extent Euler Number n
d ( Z i  Z j )
wavelet transform is continued until the sub-images reach the
Pi ¦e (7)
j 1
optimum contrast.
In (7) d is equal to 4 / r 2 where r is the neighborhood
W\ ( j , k ) h\ (n) WM ( j  1, n) (5) for each cluster. If the point falls outside this
neighborhood region it has little influence to the potential
WM ( j , k ) hM (n) WM ( j  1, n) (6)
value.
2. A point is the first center if its potential value
B. Feature extraction
P (1) is equal to the maximum of initial potential value
In this study we extract some features about the mass size,
geometrical shape and boundary after applying the FWT. The ( P (1)* ) as demonstrated in (8).
preferred features related with size are as follows (Table 1): P(1)* max( P(1)( Z i )) (8)
In a region Area is the actual scalar number of pixels,
3. A threshold ( G ) is defined for the decision to
Centroid is the center and BoundingBox is the smallest
rectangle containing the region. Filled Area is the number of continue or stop the cluster center search.
on pixels in filled image and Equiv Diameter ( 4 * Area / S ) G P u P(1) * (9)
*
is the diameter of a circle with the same area as the region. In (9) P is the reject ratio and P (1) is the potential
The features related with geometrical shape are as follows: value of the first cluster center.
Euler Number is the number of objects in the region minus 4. The previous cluster center from further
the number of holes in those objects and Extrema is the consideration is subtracted and the remaining points’
extremal points in the region. Convex Hull is the smallest potential values are adjusted using (10).
convex polygon that can contain the region. Solidity is the * 2
proportion of the pixels in the convex hull that are also in the Pi Pi  P (k ) * e  d ( Z i  Z k ) (10)
region. Finally the features related with the boundary are as *
Where Zk is the point of the kth cluster center and
follows: Major Axis Length is the length (in pixels) of the
major axis of the ellipse that has the same second-moments as *
P(k ) is its potential value. This procedure is continued
the region while Minor Axis Length is the length (in pixels) of until the maximum potential value in the current iteration
is equal to or less than the threshold G .
the minor axis of the ellipse that has the same second-
moments as the region.
Eccentricity is the eccentricity of the ellipse that has the D. ANFIS architecture
same second-moments as the region and it is the ratio of the The ANFIS is the abbreviated of adaptive neuro-fuzzy
distance between the foci of the ellipse and its major axis inference system [10]. This method is a fuzzy inference
length. Orientation means the angle (in degrees) between the system (FIS) using a backpropagation tries to minimize the
x-axis and the major axis of the ellipse that has the same Root Mean Square Error (RMSE). As in the artificial neural
second-moments as the region. Extent represents the network systems the input passes through the input layer (by
proportion of the pixels in the bounding box that are also in
the region. We also develop two further features which one is

645
Fig. 2. Sugeno fuzzy model
Fig. 3. ANFIS architecture with two inputs and an output
input membership function) and the output could be seen in
output layer. The fuzzy rules are learned by the system with
through the training process of the ANFIS. Assume that the
considered FIS has two inputs x and y and one output f
(Fig. 2). For Sugeno fuzzy model, a common rule set with
two fuzzy if–then rules is as follows:
Rule 1: If x is A1 and y is B1 , then f 1 p1x  q1 y  r1
Rule 2: If x is A2 and y is B 2 , then
f 2 p 2 x  q2 y  r 2 .
The necessary processes are implemented in five layers
as seen in Fig. 3 and the overall output is calculated.
E. Proposed algorithm
In this breast mass classification study region of interests
(ROI) are passed from a two scale FWT at first. Because it
has experienced that the sub-images become too blurred after
the second scale decomposition. We obtain 8 different
coefficient matrices using a two scale FWT.
These are approximation ( A ), horizontal ( H ), vertical
( V ) and diagonal ( D ) coefficient matrices for the first and
second scales labeled A1 , H1 ,V1 , D1 , A2 , H 2 ,V2 and D2 .
Since the first scale detail coefficients are generally
composed of poor information, only the mean of those
matrices ( M ( H 1 ,V1 , D1 ) ) is used. Consequently, six
coefficient matrices are used totally for each ROI. The next
step feature extraction is applied to those matrices separately.
646
Fig. 4. The flow chart of the proposed method.
After subtractive clustering of the feature data, ANFIS
architecture is developed. The fuzzy rules are trained
maximum 50 epochs until RMSE converges to zero. A reject
ratio of 0.5 is used and the radius is specified as 0.6. In (11)
x, y and N demonstrate targets, outputs and data size
respectively in the ANFIS.
N
¦ ( xi  y i ) 2
i 1
RMSE (11)
N
The flow chart of the proposed algorithm is illustrated in
Fig. 4.
III. THE DATASET AND EXPERIMENTAL RESULTS
The dataset which is constituted of digital grayscale
mammograms belong to the patients from Istanbul University
Faculty of Medicine Hospital in Turkey is used to test the
proposed fuzzy inference diagnosis system.
The malignant and benign masses have been pre-extracted
by a radiologist manually. Totally 78 masses (ROIs) were
obtained and 35 of them are malignant and the remains are
benign (Fig. 5). One-third of the masses (26 ROIs) were used
for testing and the others were used for training. 14 masses
are benign and the others are malignant among 26 test masses
(Table 2). Using these data sets 47 fuzzy rules are extracted
in the whole test process. Before the ANFIS training process
the initial error was 0.14. As seen in Table 3, when the
network is trained 50 epochs which produce best
TABLE 2
Number of the masses in training and test sets
Train set Test set
Malignant 23 12
Benign 29 14
 IV. CONCLUSIONS
Breast cancer disease is among the most serious cancer
types in the world but curable ones if it can be diagnosed
early. Computer aided diagnosis systems used for medical
decision provide medical data to be examined in shorter time
and in more detail. The research presented in this article aims
to decrease the mortality rate related to breast cancer by
reducing the number of malignant masses which radiologists
may miss by means of computer aided techniques. We
developed a program in MATLAB 7.6 combined with the
methods FWT and ANFIS which classifies the pre-extracted
ROIs as malignant or benign. One can see that this method is
efficient for solving the real world problems related with
breast cancer diagnosis using FWT multi-resolution
Fig. 5. Some of the ROIs in different characteristics: (a), (b) malignant
decomposition and ANFIS rule extraction and learning
masses (c) benign masses.
methods. The satisfying performance of 92% classification
TABLE 3 rate demonstrates that this study is valuable to improve early
Experimental results of the FWT-ANFIS method. (Tr.Ep. : Training epoch, diagnosis and reduce the number of unnecessary biopsies.
Tr. Err. : Training error, Az. : Area under the ROC curve, Sen.(Rec.) :
Sensitivity (Recall), Spe. : Specificity, Pre. : Precision, FMs. : FMeasure,
Acc. %: Accuracy) REFERENCES
Tr. Tr. Az Sen. Spe. Pre. FMs. Acc. [1] S. Sahan, K. Polat, H. Kodaz, S. Gunes, “A new hybrid method based
Ep. Err. (Rec.) % on fuzzy-artificial immune system and k-nn algorithm for breast cancer
20 0.041 0.88 75 100 100 86 88 diagnosis”, Computers in Biology and Medicine, vol. 37, pp. 415 –
30 0.032 0.88 83 93 91 87 88 423, 2007.
40 0.024 0.89 100 79 80 89 88 [2] P. Delogu, M. E. Fantacci, P. Kasae, A. Retico, “Characterization of
50 0.012 0.93 100 86 86 92 92 mammographic masses using a gradient-based segmentation algorithm
60 0.020 0.89 100 79 80 89 88 and a neural classifier”, Computers in Biology and Medicine, vol. 37,
70 0.028 0.89 92 86 85 88 88 pp. 1479 – 1491, 2007.
80 0.035 0.88 83 93 91 87 88 [3] R. M. Rangayyan, N. R. Mudigonda, J. E. L. Desautels, “Boundary
modelling and shape analysis methods for classification of
TABLE 4 mammographic masses”, Medical and Biological Engineering and
Confusion matrix. (Tr.Ep. : Training epoch, TP: True Positives, TN: Computing, vol. 38, no.5, pp. 487-496, 2000.
True Negatives, FP: False Positives, FN: False Negatives) [4] D. Cascio, F. Fauci, R. Magro, G. Raso, “Mammogram segmentation
Tr. Tr. Tr. Tr. Tr. Tr. Tr. by contour searching and mass lesions classification with neural
Ep. Ep. Ep. Ep. Ep. Ep. Ep. network ”, IEEE Transactions on Nuclear Science, vol.53, no.5, pp.
20 30 40 50 60 70 80 2827-2833, 2006.
TP 9 10 12 12 12 11 10 [5] D. Tralic, J. Bozek, S. Grgic, "Shape analysis and classification of
TN 14 13 11 12 11 12 13 masses in mammographic images using neural networks", 18th
FP 0 1 3 2 3 2 1 International Conference on Systems, Signals and Image Processing
FN 3 2 0 0 0 1 2 (IWSSIP), Sarajevo, 2011.
[6] S. Mallat, “A compact multiresolution representation: The wavelet
model”, IEEE Computer Society Workshop on Computer Vision,
classification, the training error is decreased to 0.012 and the Washington, 1987.
classification accuracy reaches to 92%. If the epoch number [7] R. C. Gonzales, R. E. Woods, Digital Image Processing using
MATLAB, Prentice Hall, USA, 2004.
is increased or decreased, the error changes but the accuracy [8] M. Alata, M. Molhim, “Optimizing of fuzzy C-means clustering
becomes stable at 88% each time. Algorithm using GA”, World Academy of Science, Engineering and
In Table 4 the confusion matrices are given obtained after Technology, vol. 39, 2008.
different training epochs. To emphasize the significance of [9] Elmzabi, M. Bellafkih, M. Ramdani, “An adaptive fuzzy clustering
approach for the network management”, World Academy of Science,
the FWT, the system is tested without FWT and the best
Engineering and Technology, vol.31, pp. 425-430, 2007.
classification performance of 85% is calculated. As known, [10] P. Tahmasebi, A. Hezarkhani, “Application of Adaptive Neuro-Fuzzy
the higher the performance is, the more accurate the diagnosis Inference System for Grade Estimation; Case Study”, Australian
is. Journal of Basic and Applied Sciences, vol.4, no.3, pp. 408-420, 2010.

647