Content uploaded by Ayush Singh
Author content
All content in this area was uploaded by Ayush Singh on Jan 14, 2018
Content may be subject to copyright.
Classification of Hematomas in Brain CT Images
using Support Vector Machine
Devesh Kumar Srivastava 1, Bhavna Sharma 2, Ayush Singh3
1, 3 SCIT, Manipal University Jaipur, Rajasthan, India
2Apex Institute of Engineering & Technology, Jaipur, India.
Email: 1 devesh988@yahoo.com, 2bhavnapareek@rediffmail.com
3mail.ayushsingh@gmail.com
Abstract. Hematoma is caused due to traumatic brain injuries. Automatic
detection and classification system can assist the doctors for analyzing the brain
images. This paper classifies the three types of hematomas in brain CT scan
images using Support vector machine (SVM). The SVM has been simulated and
trained according to the dataset. Trained SVM classifiers performances were
compared on the basis of parameters i.e. classification accuracy, mean square
error, training time and testing time. The classification process depends on the
training dataset and results are based on simulation of the classifiers.
Keywords: CT scan, Hematoma, Data Mining, Support Vector Machine.
1 Introduction
Brain Hematoma is caused due to traumatic injury in brain and blood is collected
inside the brain [1]. Three types of brain hematomas are Epidural Hematoma (EDH),
Subdural Hematoma (SDH) and Intracerebral Hematoma (ICH). Hematomas are
hyper dense in nature so they appear brighter than the other brain tissues. EDH
involve bleeding between the skull and the dura matter of the brain and has a
biconvex shape while a SDH is caused between dura and subarchnoid layer of the
brain having a crescent shape. ICH is inside the brain having a circular shape.
The CT scan is among various available brain scanning techniques, it is preferred in
comparison to the other techniques because of its extensive availability, little cost, fast
scanning and superior contrast [2]. Classification task is the most frequently
encountered problem in medical image analysis. In these tasks an object should be
allocated into a particular group or class depending upon the observed properties or
attributes called features related to that object. Many classifiers are available,
developed and successful in medical imaging applications but all are having different
initial parameter that varies according to the applications [3]. The accuracy of
classification methods must be high because the diagnosis and treatment is based on
this categorization. This paper focuses on SVM classifier for classifying the
hematomas in brain CT images. There are varieties of other approaches available for
such an application [4]. There are no clear rules or procedures that can be followed to
choose the best classifier for this specific problem [5]. We compared performances of
classifiers on the basis of parameters i.e. classification rate, mean square error,
training and testing time etc. The classification process depends on the training
dataset and results are based on simulation of the classifiers.
2 Problem Statement
The aim of the research paper is classification of brain Hematomas by extracting
features from brain CT images.. Figure 1 shows axial slice of brain CT scan images
for normal and three types of hematomas i.e. Epidural (EDH), Intracerebral (ICH) and
Subdural (SDH) respectively. The organization of the research paper consists of
introduction and problem statement given in section I and II respectively, and section
III describes methodology for brain CT image analysis. Section IV describes the
simulation and training of two SVM classifiers (one against one and one against all)
and interprets the result. Lastly, we conclude the paper and discuss the future work.
Figure 1. Normal Image and EDH Image
Figure 2. CT Scan images of Brain with
ICH and SDH
3 Methodology
The Methodology for analyzing the brain CT Image is divided into several steps as
shown in Figure 2. In the first step image is acquired from CT scan machine that is
converted into suitable format. We collected images of brain CT scan in DICOMM
and convert it into jpeg format. The images are from same CT scan machine and
having resolution of 512*512 with same window width and window level. Second
step is to preprocess the image to remove noise from the image. The brain portion is
extracted to remove the artifacts from the image. The extracted region contains only
the brain portion, which is further processed in analysis steps. Third step segments the
image [6] automatically into regions. A new hybrid method [7] based on peak
detection and K-means clustering is used, that automatically segments the
preprocessed image as shown in Figure 3.
Figure 3. Steps in brain CT image analysis for hematomas classification
In the fourth step statistical features [8], intensity features [9], shape features [10]
and texture features [11] are extracted from the segmented image. These features are
obtained from each tested image. Large number of features would be more
informative but due to the complexity, computational cost, increase in generalization
error, feature selection [12] methods are used that bring out the most informative
features. We used backward feature selection method in which we started with all
feature vectors and then remove the features one by one, at each step, until any further
removal of the feature increases the error (misclassification rate) significantly. After
rigorous checking and many trials, finally 12 features are selected these are given in
Table 1. Area of the segmented region is 3055 pixels that is important in calculating
the size of hematomas and other important features but not used as an independent
feature. In last step, classification task which is organizing the data into different
groups on the basis of these features consists of training and testing phase. In training
phase, features of the image are presented to the network and a unique description of
each class is created. In testing phase, these features are used to classify images in
categories. Multi-class SVM (one-against-one (O-V-O) and one-against-all (O-V-A))
has been simulated and trained [13] according to the training samples presented to the
system. Performances of each classifier are evaluated according to the performance
parameters.
Figure 4. Preprocessed image.
Figure 5. Original image and segmented image.
Table1. Extracted Features and their values obtained from the image in Figure
Segmented image
S. No.
Feature
Value(in pixels)
1
Circularity
0.0830
2
Major Axis
235.18
3
Minor axis
37.34
4
Eccentricity
0.9834
5
Solidity
0.3245
6
Extent
0.1008
7
Convex Hull
Area
10564
8
Adjacent to
skull boundary
Yes(1)
9
Ratio of
bounding box
width to minor
axis
0.2534
10
Energy(ASM)
0.9621
11
Entropy
0.1188
12
Contrast
0.1559
4 Experimental Results
The entire methods are implemented in MATLAB 7.1 on Windows 7. Computer
configuration consists of i3 processor, 4 GB of RAM and 2GB graphics Card. The
experiment data consists of 150 CT images of brain. The images are of same quality
but with different resolutions. The 50 images are of normal brain and rest has
hematomas of different types, having different shapes and at different locations. We
selected 12 features among several features from segmented images. These 12
features from each image with a sample size of 100 are presented to classifiers as
input patterns and 3 outputs are generated for three types of hematoma i.e. EDH,
SDH, ICH. The samples are randomly divided into training, validation and testing
sets.
SVM [14] is supervised binary classifier capable of giving high performance in
classification of patterns. Although SVMs are designed to do only two-class
classification, they have shown consistent performance when applied to multi-class
classification problems, using one-against-one (O-V-O) or one-against-all (O-V-A)
approaches. The decision making with the O-V-O strategy is more complex than with
the O-V-A and the time taken to train c(c-1)/2 is lesser than training with c SVM’s(c
represents the no. of classes).The complexity of an SVM’s decision making is directly
linked to the number of support vectors and although the decision making is more
complicated in the multi-class case, it is reasonable to consider that the complexity is
proportional to the total number of support vectors. However, in our experiments the
O-V-O strategy uses less support vectors than the O-V-A. We checked classification
accuracy by SVM with different kernel functions against various combinations of
training data set. Data is divided randomly in training and testing sets with different
proportions. We found that RBF kernel function gives better classification
performance than the other kernels in both types of SVM as shown in Figure 5 and
Figure 6.
SVM scales well and the training of data is relatively easy. Best suited parameters and
model can be chosen for the task. From the Figure 5 and 6, it is clear that as the
percentage of training data increases accuracy level increases. So training with more
than 60% of data gives more classification accuracy. RBF kernel function qualifies
among other kernels in both SVMs and accuracy of classification is better in O-V-O
SVM than O-V-A SVM.
Random division of the data into training and testing set and running it many times
gives different accuracy levels. Moreover, the random division might bring about
training/test datasets with different proportions of output classes. Even if we tune all
the parameters according to the model that gives highest classification accuracy this
might not result in a reliable evaluation of the network.
An improvement to the random sampling is cross validation. In cross validation, each
record is used the same number of times for training and exactly once for testing. If
data is partitioned in two equal sized subsets, one of the subset for training and the
other for testing then the roles of the two subsets are swapped. This approach is called
a twofold cross validation. The total error is obtained by summing the errors for both
the runs. In this paper we used 10 fold cross validation where data set is divided into
10 folds randomly and the classifier is tested with one fold and trained with remaining
folds. This process repeated 10 times such that data is tested with each fold. Finally
average of training, testing and overall accuracy is calculated.
Figure 6. Classification accuracy in O-V-A multiclass SVM with linear, RBF and
polynomial kernel functions with varying number of data set size for training and
remaining data for testing chosen randomly.
Figure 7. Classification accuracy in O-V-O multiclass SVM with linear, RBF and
polynomial kernel functions with varying number of data set size for training and
remaining data for testing chosen randomly.
In SVM classifier we apply the 10 fold cross validation for both one against one (O-
V-O) and one against all (O-V-A) and the training and testing accuracy are measured
[15]. In SVM RBF kernel function is used as it is more reliable than other kernel
functions. We considered the RBF kernel and used 10 fold cross-validation to find the
best parameter C and γ. Pairs of (C, γ) are tried and the one with the best cross-
validation accuracy is picked. Trying exponentially growing sequences of C and γ is a
practical method to identify good parameters. After rigorous checking and running
several times we choose (C, γ) as (1.0, 0.5).
After selecting suitable parameters the testing accuracy on 10-folds are calculated and
the results proved that the O-V-O SVM gives the average testing accuracy of 97%
which is higher than the O-V-A SVM having 88% of average training accuracy.
Applying 10 fold cross validation on each classifier testing accuracy varies with each
fold. This may be due to the fact that features are randomly partitioned in each fold so
calculation of average accuracy is necessary. O-V-O multi class SVM gives average
accuracy of 97% and O-V-A multi class SVM gives average accuracy of 88%.
The important factor in simulating these classifiers are training time and setting the
right initial parameters which is again a time consuming task. We ignored here
parameters selection time. The training time and testing time for classifiers after
setting the initial parameters are shown in Table 2. This is the time of testing one fold
and training given by 9 folds. There is not much variation in training time and testing
time for each fold so we have not shown the time for each fold. Average training and
testing time are shown in Table 2, we conclude that training time for O-V-O
multiclass SVM classifier is lesser than O-V-A multiclass SVM.
Table2. Comparison of testing and training time of SVM classifiers
Classifiers
Training time
Testing time
SVM(O-V-O)
0.015600
0.00015
SVM(O-V-A)
0.043300
0.0013
Location is also an important factor in the surgical decision. If temporal hematomas
are large or expanding that may lead to herniation and more rapid deterioration. EDH
in the posterior fossa, often requires prompt evacuation because of the limited space
available compared with the EDH on frontal compartment. In this system for finding
the location of the hematomas we have divided the image in four segments and find
out the pixels in the ROI belongs to which segments and at what ratio.
We only considered four lobes of the brain and set rules according to expert
knowledge to find the location of the hematomas. The first segment is right frontal
lobe, second segment is left frontal lobe and third and fourth segment belongs to
parietal and temporal lobe and if hematoma is in third and fourth segment than it is in
posteria fossa . This whole procedure takes 0.5 seconds on an average as the
calculation that how many pixels belongs to which segments needs labeling and
counting the pixels in each segment. Here we have found the approximate location of
the hematomas but for finding the exact location for example near the ventricles, syri,
CSF etc. more segments are needed, more segmentation of the brain axial slice should
be done and more knowledge should be incorporated into the system.
For calculating the size of hematomas, area of region feature extracted from
segmented image of Figure 9 is considered. As the radiologist are concerned with not
only the size of hematoma region but also the width of the region. We have calculated
the width of EDH and ICH from the minor axis of the fitted ellipse on the region and
multiplying with the size of pixel. Due to SDH shape, fitted ellipse minor axis has
almost double than the region so width of SDH region is calculated by dividing minor
axis by 2 and multiplying with pixel size.
The Pixel size is calculated from field of view (FOV) and by the matrix size, for
example if CT FOV is d and the matrix size is M, then for a typical head scan with an
FOV of 10 cm and a matrix of 512 pixels the pixel size is 0.2 mm.
Pixel size =d (in mm)/M = 100/512 = 0.2 mm
Figure 8. Division of image into four equal parts (a) Original which shows that SDH
is located in Left-Frontal-Parietal lobe (b) Percentage of pixels in each segment in
segmented image
In practice, the pixel size is smaller than the theoretical values due to limiting
resolution imposed by the finite sizes of the focal spot and x-ray detectors [19]. Axial
resolution within the scan plane can be improved by operating in a high resolution
mode using a smaller FOV or a larger matrix size. For the image in Figure 9, Area of
the region is 3055 pixels and minor axis of fitted ellipse is 38 pixels than the size and
thickness of the SDH is calculated as
Size of SDH = (Area)* Pixel size = 3055 *0.2 = 611mm2
Thickness of SDH is = (Minor axis)/2 * Pixel Size =17 * 0.2= 3.4mm
This is moderate size SDH and surgical evacuation via craniotomy is often considered
in patients with a SDH thicker than 10 mm. There is no problem in recommending
surgery in the large size hematomas and recommending medicines and therapies in
smaller size hematomas. The main problems are with moderate size hematomas that
requires expert knowledge and other parameters like age, Glasgow comma scale. The
brain image analysis and recommendation for diagnosis may vary with radiologist.
5 Conclusion
In this paper SVM classifiers are proposed to classify the hematoma types in brain CT
images. The input dataset (features) to train the classifiers are obtained from
segmented CT scanned images. Constructing and setting initial parameters in SVM
classifier is simpler than other classifiers. SVM require less trial and error and thus,
less time and effort for better getting classification rate. Overall performance of SVM
(O-V-O) classifier outperforms others in terms of complexity, lesser number of
training samples, lesser time, effort and high classification accuracy.
Our Future work is to find out the exact location and exact volume of hematomas by
adding more expert knowledge in the system. However, the ultimate future goal of
this ongoing research is to provide an everyday clinical tool that will assist clinicians
in the diagnosis of brain hematomas. To achieve this system must undergo exhaustive
clinical trials to ensure its proper functionality.
References
1. Atam P. Dhawan, H. K. Huang, Dae Shik Kim 2008. Principles and
Advanced Methods in Medical Imaging and Image Analysis. World Scientific.
2. D. Seletchi, O. G. Duliu, 2007.Image processing and data analysis in
computed tomography. Roman Journal of Physics, Vol. 52, No. 5, 2007, pp. 667-
675.
3. C.J.C. Burges, 2002.A Tutorial on Support Vector Machines for Pattern
Recognition. Data Mining and Knowledge Discover, Vol. 2, No. 2, pp. 121-167.
4. Cosic, S. Loncaric, 1997.Rule-based labeling of ct head image. Proceedings
of the sixth Conference on Artificial Intelligence in Medicine in Europe (AIME
1997), pp. 453-456.
5. W.L. Zhang, X. Z. Wang, 2007 “Extraction and classification for human
brain CT images”, Proceedings of Sixth International Conference on Machine
Learning and Cybernetics, IEEE, pp 1155-1156.
6. B. Sharma, K. Venugopalan, 2012.Automatic segmentation of brain CT
scan images. International Journal of Computer Applications, Vol. 40,No.10, pp. 1-
4.
7. B. Sharma, K. Venugopalan, 2012.Classification of hematomas in brain CT
images using neural network. Proceedings of IEEE Conference on Issues and
Challenges in Intelligent Computing Techniques, pp. 41-46.
8. Nixon, Mark, and Alberto S. Aguado 2008. Feature Extraction & Image
Processing. Academic Press.
9. Yang, M., Kidiyo K. and Joseph R. 2008. A survey of shape feature
extraction techniques. Pattern Recognition, pp. 43-90.
10. Haralick, Robert M., Karthikeyan S., and Its Hak Dinstein 1973.Textural
features for image classification. IEEE Transactions on Systems, Man and
Cybernetics, 6, pp. 610-621.
11. Kohavi R. and John G. H. 1994.Wrappers for feature subset selection.
Artificial Intelligence Vol. 97, No. 1, pp. 273-324.
12. Gong, T., Li, S., Wang, J., Tan, C. L., Pang, B. C., Lim, C. T., & Zhang, Z.
2011.Automatic labeling and classification of brain CT images. 18th IEEE
International Conference on Image Processing (ICIP),pp. 1581-1584.
13. Vapnik 1999. Support-Vector Networks. Machine Learning, pp. 273-297.
14. J.A.K. Suykens, T. Van Gestel, J. De Brabanter, B. De Moor, J. Vandewalle
2002. Least Squares Support Vector Machine. World Scientific Publishing Co.,
Singapore.
15. Chang and C. Lin. 2011 .LIBSVM: a library for support vector machines.
ACM Transactions on Intelligent Systems and Technology, Vol. 2, No. 27, pp. 1-
27. Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm.
