Beruflich Dokumente
Kultur Dokumente
Abstract - X-Ray is one the oldest and frequently used presentation and their outcomes are unpredictable. Tibia
devices, that makes images of any bone in the body, fractures are the most common long bone fracture accounting
including the hand, wrist, arm, elbow, shoulder, foot, to more than 20 per cent occupancy of hospital wards [9].
ankle, leg (shin), knee, thigh, hip, pelvis or spine. A According to [27] , on average 26 tibia fractures occur per
typical bone ailment is the fracture, which occurs when 1,00,000 populations per year.
bone cannot withstand outside force like direct blows,
twisting injuries and falls. Automatic detection of The numerous incidences necessitate the healthcare
fractures in bone x-ray images is considered important, as professionals to analyze huge number of x-ray images. As a
humans are prone to miss-diagnosis. The main focus of consequence of continuously analyzing images, cases of miss
this paper is to automatically detect fractures in long diagnosis occur. A miss is defmed as either failure to see a
bones and in particular, leg bone (often referred as significant fmding, or attaching the incorrect significance to a
Tibia), from plain diagnostic X-rays using a multiple fmding that is readily seen. A high miss-diagnosis rate will
classification system. Two types of features (texture and result in poor quality in healthcare and time-delayed
shape) with three types of classifiers (Back Propagation treatment and has to be avoided. All these facts have
Neural Network, K-Nearest Neightbour, Support Vector motivated us to focus on the proposed fracture identification
Machine) are used during the design of multiple and detection system that can automatically determine the
classifiers. A total of 12 ensemble models are proposed. presence and absence of fractures and if present, methods that
Experiments proved that ensemble models significantly could locate the fractures in a bone x-ray image.
improve the quality of fracture identification. Classification, a frequently used data mining technique, has
been used widely to detect the presence of fracture for the
I. INTRODUCTION past few decades [2S]. These systems combine various
features (like shape, texture and colour) extracted from x-ray
In today's revolution oriented medical environment, images and machine learning algorithms to identify faults.
computer aided disease detection play a vital role in a wide
range of applications and services in day-to-day activities. According to [32] when a perfect set of features that can
The most important quality goal of such a system is speed describe the image data is given, the accuracy of the resultant
and high accuracy in disease detection. This is more classification depends on the classifier adopted. Several
important in radiology department, where hundreds and solutions have been proposed for this purpose. Among which,
hundreds of images are to be examined for various ailments the usage of Neural Network (NN), K Nearest Neighbour
including bone fracture, infection and location of foreign (KNN) and Support Vector Machines (SVM) based
objects. Out of this, bone fracture detection is the most classifiers are more prominent. However, as pointed out by
frequently used task, because the population affected by this [30], the success rate of a classification problem can be
ailment has increased due to the increased reports of improved by using multiple classifiers and then the various
accidents and due to osteoporosis. According to the market results can be fused together to obtain an optimum
report of [ IS], the fracture features are more in Indian classification. Using multiple classifiers (either different types
hospital records and the incidence has increased 3-folds over of classifiers or different instantiations of the same classifier)
the past 35 decades with more than 4.4 lakh people and is improve the success rate of an classification model. This
expected to increase to more than 6 lakh in 2020 [21]. process is termed as 'ensemble classification' or 'fusion
Among fractures, automatic detection in tibia is considered classification' or 'multiple classification system'. According
more chaUenging because they are different and variable in to [31], intuitively, fusion classification allows the different
ICCCNT'12
26 th_2Sth July 2012, Coimbatore, India
IEEE-20180
Classifier 1
Fracture Present
Features
X-Ray Image Classifier 2
Fracture Absent
Classifier 3
L ________________________
ICCCNT'12
26 th_28th July 2012, Coimbatore, India
IEEE-20 ISO
and having the same size as the original training data. The preprocessing step enhances the x-ray images through the
Multiple classification is obtained by means of uniform use of a hybrid denoising method that uses Independent
majority voting, where an unlabeled observation is assigned Component Analysis (lCA) coupled with wavelets. The
the class with the highest number of votes among the implementation details are given in [37]. The second step,
individual classifiers' predictions. Theoretically, segmentation is used to extract the bone structure from the x
bootstrapping can induce large differences in the constructed ray images. For this purpose, an enhanced active contour
individual classifiers which substantially improve the model that uses region growing algorithm to estimate the
accuracy of the fusion classifier. Several variations upon initial seeds is used. Further, as the region of interest to detect
Bagging have been proposed in search for further fractures is considered as the diaphysis of tibia, a
performance improvements. Two popular strategies involve segmentation method that extracts only the diaphysis region
(i) increasing variation in the training data for base classifiers is proposed. This method uses a fast hough transformation
and (ii) the use of alternative base classifier algorithms. method. Detailed description of the segmentation process is
Firstly, several studies have shown the impact of variations presented in [39]. General details on tibia and diaphysis can
of the input data used for the training of base classifiers. be found in previous published in [3S].
Varying the training data of the members of an fusion
classifier is a strategy to increase diversity amongst member The feature extraction stage is used to extract various features
class ifiers, which is generally perceived as a key driver of that best exhibit the characteristics of the segmented image.
fusion performance [26]. In the Random Subspace Method Twelve features, which can be grouped into texture features
(RSM), variables are randomly sampled to create training and shape features, are used to generate the feature vector.
data sets for a decision tree classification fusion. RSM, also The texture features collected are GLCM (Gray Level Co
referred to as Attribute Bagging [6], specifies that each Occurrence Matrix) features, namely, Contrast, Homogeneity,
fusion member is trained using a random feature subset Energy, Entropy, Mean, Variance, Standard Deviation,
(RFS), i.e. a random selection of explanatory variables Correlation, Gabor orientation (GO), Markov Random Field
sampled without replacement and of a predefmed size. (MRF), and intensity gradient direction (IGD). The shape
features are extracted using a Fast Hough Transformation
A related method is the Random Forest algorithm by [3], proposed by [16]. The features collected are arranged in a
which has demonstrated high classification performance in two-dimensional matrix where each column represents a
many fields of research [1]. A Random Forest combines feature extracted and the row represents various features that
Bagging and a specific form of RSM where random feature represent the segmented x-ray image. The final column is
subset selection is performed at each node of a member treated as a target label column, having a Boolean value of 0
decision tree. Rotation Forest [34], a multiple classifier based to indicate the absence of fracture and 1 to indicate the
on rotations of the feature space through Principal presence of fracture. Us ing these 12 features, three feature set
Component Analysis (PCA). The purpose of Rotation Forest vectors are created. The frrst one consists of only texture
is to increase the individual classifier performance and the features, second have only shape features and the third set
diversity within the multiple classification process. Diversity have a shape and texture features and are referred to as FS I,
is achieved for each classifier by applying feature extraction, FS2 and FS I2 respectively in this paper.
while one tries to increase the performance by using all
principal components and training the model on the whole The present study considers three classifiers namely, Support
data set. Vector Machine (SVM) [15], K-Nearest Neighbour (KNN)
[10] and Back Propagation Artificial Neural Network
A second strategy to increase classification performance is to (BPNN) [7]. All the classifiers are designed as binary
select an alternative base classifier algorithm. Many studies classifiers, where the classification result is either 'Fracture
have proposed fusion based on alternative base classifiers, Present' or 'Fracture Absent'. During experimentation, a
such as Artificial Neural Networks [40], Support Vector standard three-layered back-propagation network with the
Machines [22], parametric regression techniques [33] and tangent-sigmoid transfer function is considered. The weights
nonparametric regression techniques [2]. and biases of the neural networks are initialized randomly,
and the number of neurons in the hidden node is determined
III. PROPOSED CLASSIFICATION SYSTEM heuristically as inputs + outputs. A small value of the learning
rate (0.15) and a large value of the momentum rate (O.S) are
The proposed ensemble classification system consists of six chosen to avoid local minima. The number of training epochs
steps. was 500. To implement the principles of SVMs, we used the
LIB-SVM. The two most important steps in implementation
l. Preprocessing of SVM is scaling and kernel selection; for scaling, the values
2. Segmentation of all features were linearly scaled to the range [1, + 1] to
3. Feature Extraction and Normalization prevent the cases that features great numeric ranges
4. Training of classifiers dominating those in smaller numeric ranges. Among many
5. Assess classifiers with respect to diversity available kernel functions linear kernel was used. The KNN
6. Perform class ification fusion algorithm was constructed using k=2.
ICCCNT'12
26 th_2Sth July 2012, Coimbatore, India
IEEE-20 IS0
The classifiers are trained using boosting [14] method where The fusion method used was majority voting method.
different feature subsets are used to train the classifiers. The Majority voting scheme is one of the oldest strategies used to
accuracy of the single classifiers is evaluated using lO-fold aggregate the results of multiple classifiers in an ensemble
cross validation technique. Table I shows the performance of system. It is one of the simplest and fastest algorithm and the
the three classifiers with respect to classification accuracy. method is explained as follows. Let the decision of the i th
classifier be defmed as dt,j E {O, I}, t I, . . . , T and j I, . . . ,
= =
Table 1 : Ten-fold Validation Classification Accuracy C, where T is the number of classifiers and C is the number of
classes. If the ith classifier chooses class (OJ, then dt,j I and 0,
=
T c T
(2)
L dt.J = max L dt.j
The results from the above table show that while considering t=l J=l t=l
accuracy the classifiers are roughly comparable. As the next
step, the diversity of the classifiers is considered and are With majority voting the fusion classifier makes the correct
estimated using Yule's Q method [23], where Q between two
decision if at least L TI2J + I classifiers choose the correct
classifiers is calculated using the confusion matrix (Figure 2)
label, where the floor function L J returns the largest integer
.
= ad- bc (1)
Q IV. EXPERIMENTAL RESULTS
ad+bc
The experimental results of ensemble classification with
Classifier I Classifier I different base classifiers are presented in the following
(Correct) (Incorrect) sections. To analyze the performance of the ensemble systems
Classifier 2 A b the commonly used measures, namely Accuracy, Precision,
(Correct) Recall and F-Measure are used. All the measures are
Classifier 2 C d calculated from the confusion matrix (Figure 2). Accuracy is
(Incorrect) a metric that is used to measure the correct classification rate
and is defmed as the ratio of number of images correctly
Figure 2 : Confusion Matrix predicted to the total number of images (Equation 3). The
second measure, precision, also used to analyze the
{<
The resultant Q value ranges between -I to I and can be correctness of a classifier is the ratio of the number of images
interpreted as follows: corrected classified as having fracture to the total number of
images predicted as having fracture (Equation 4). The recall,
0 Uncorrelated ClassifielS otherwise known as defect detection rate, is the ratio of
number of images corrected classified as having fracture to
= 0 Statistically Independent ClassifielS the total number of images that actually have fracture
(Equation 5). F Measure an amalgamation method of
> 0 CorrelatedClassifielS
precision and recall presents the accuracy in a more precise
manner and can be calculated using Equation (6).
Table 2 shows the Q values obtained for the three selected
classifiers.
a+d (3)
Accuracy= - -:-- --,
Table 2 Q Values for the Selected Classifiers
-
: a+b+c+d
d (4)
Precislon=
.. -
Precision+Recall
From the table, it could be seen that even though all
algorithms are moderately correlated, they exhibit high To analyze the effectiveness of the ensemble classifiers,
pairwise diversities to each other. Thus, it could be seen that statistical analysis of the results was conducted. 't' test
the selected base classifiers have high diversity and can be method proposed by [29] was used for this purpose. The NB
used to design ensemble system. The present study considers 't' test was performed at 95% confidence level (0.05 level) to
two-classifier and three-class ifier ensemble systems. The analyze the significant difference between the single
classifiers were built while varying the feature sets. Thus, a classifiers and ensemble classifiers. The NB-'t' test method
total of 12 unique combinations of classifiers using three was adopted because it is more suited for classifiers adapting
single classifiers and 3 feature sets were designed. lO-fold cross-validation method [12]. The traditional student
ICCCNT'12
26 th_2Sth July 2012, Coimbatore, India
IEEE-201S0
V.CONCLUSION
ICCCNT'12
26 th_2Sth July 2012, Coimbatore, India
IEEE-201S0
ICCCNT'12
26 th_2Sth July 2012, Coimbatore, India
IEEE-20180
ICCCNT'12
26 th_28th July 2012, Coimbatore, India
IEEE-20180
[26] L.I. Kuncheva, and C.J. Whitaker, [39] N. Umadevi and Dr. S. N. Geethalakshmi,
Measures of diversity in classifier ensembles Enhanced Segmentation Method for Bone
and their relationship with the ensemble Structure and Diaphysis Extraction from X
accuracy, Machine Learning, Vol. 51, No.2, Ray Images, IJCA, Vol.37,No.3, Pp. 30-36,
Pp. 181-207, 2003. January 2012.
[27] S.E. Lim, Y. Xing, Y. Chen, W.K. Leow, [40] Z. H. Zhou, J. X. Wu, and W. Tang,
T.S. Howe, and M.A. Png, Detection of Ensembling neural networks: Many could be
femur and radius fractures in x-ray images, better than all. Artificial Intelligence, Vo I.
2nd International conference on Advances in 137, Issues 1-2, Pp. 239-63, 2002.
Medical Signal and Information Processing,
Malta G.C, 2004.
[28] M.E. Muller, S. Nazarian, P. Koch, and J. AUTHORS BIBLIOGRAPHY
Schatzker, The comprehensive class ification
of fractures of long bones. Berlin: Springer
Verlag, 1990. Dr.S.N.Geethalakshmi is working as Associate
[29] C. Nadeau, and Y. Bengio, Inference for the Professor in the Department of Computer
generalization error, Machine Learning, Vol. Science in Avinashiligam Institute for Home
Science and Higher Education for Women,
52, Pp.239-281, 2003.
Coimbatore. She has more than 20 years of
[30] N.V. Neeba and C.V. Jawahar Empirical
teaching experience. Her research interest
evaluation of character classification includes Image Processing and Software
schemes, Seventh International Conference Engineering. She has around 50 publications in
on Advances in Pattern Recognition her research area at National and International
(lCAPR), Pp. 310-313, 2009. Level. Presently she is guiding M.Phil and
[31] N.C. Oza, and K. Turner, Classifier Ph.D research scholars. She is currently the
ensembles: Select real-world applications, Principal Investigator of one of the Major
Journal of Information Fusion, Vol. 9, Issue research Project funded by NRB. She is also a
life member of one of the Professional
1, Pp. 4-20, 2008.
Organization in the Indian Science Congress
[32] D.C. Park, Image Classification Using
Association.
Partitioned-Feature based Classifier Model,
International Conference on Computer
Systems and Applications (AICCSA),
IEEE/ACS, Pp.l-6, 2010.
[33] A. Prinzie, and D. Van den Poel, Random
forests for multiclass classification: Random N.Umadevi working as Head in the
MultiNomial Logit. Expert Systems with Department of Computer Science and
Applications, Vol. 34, No. 3, Pp. 1721-32, Information Technology cum NSS
2008. Programme officer, Sri Jayendra Saraswathy
[34] J.J. Rodriguez, L.I. Kuncheva, and C.J. Maha Vidyalaya College of Arts and Science,
Singanallur,Coimbatore is a part time
Alonso, Rotation forest: A new classifier
Research Scholar of Avinashilingam
ensemble method. IEEE Transactions on University for Women, Coimbatore doing her
Pattern Analysis and Machine Intelligence, research under the guidance of
Vol. 28, No.lO, Pp. 1619-1630, 2006. Dr.S.N.Geethalakshmi. The Scholar has 3
[35] H. Schwenk, and Y. Bengio, Boosting neural years of industrial experience and 10 years of
networks, Neural Computation, Vo1.12, No.8, teaching experience. Her area of interest are
Pp. 1869-1887, 2000. Image Processing and Data Mining. Her
[36] M. Skurichina, and R.P.W.Duin, The role of publications include 5 International Journals,
combining rules in bagging and boosting, 3 International Conferences and 3 National
Conferences.
2000, F.J. Ferri, J.M. Inesta, A. Amin, and
P. Pudil, (Eds.), Proc. of Joint International
Workshops SSPR 2000 and SPR 2001,
Springer-Verlag, Berlin I Heidelberg,
[37] N. Umadevi and Dr. S. N. Geethalakshmi,
Improved Hybrid Model for Denoising
Poisson Corrupted X-Ray Images, IJCSE,
Vol.3,Issue 7, Pp. 2610-2619, July 2011.
[38] N. Umadevi and Dr. S. N. Geethalakshmi, A
Brief Study on Human Bone Anatomy and
Bone Fractures, IJCES, Vol.l,Issue 3,Pp. 93-
104, December 2011
ICCCNT'12
26 th_28th July 2012, Coimbatore, India