International Journal of Advanced Information Science and Technology (IJAIST) ISSN: 2319:2682

Vol.5, No.5, May 2016 DOI:10.15693/ijaist.2016.v5.i5.61-64

Lung Cancer Detection Using Bayesian

Classifier
Hitha Rocky Dr.Jereesh A.S
Department of Computer Science and Engineering Assistant professor/Department of Computer Science and
Adi Shankara Institute of Engineering and Technology, Engineering
Kalady Cochin University of Science and Technology,
Ernakulam

Abstract— Medical image processing is being widely used in the

medical field for early detection of chronic diseases like lung
cancer. Time is an important factor in the treatment of lung
cancer. . If lung cancer is detected in time, the overall 5-year
survival rate of cancer patients increases from 14 to 49% .
Computed Tomography (CT) images being more efficient than
X-rays are preferred for diagnosis. The challenge lies in choosing
the best/most accurate segmentation and classification technique
for isolating the cancer affected parts. Firstly, segmentation is
used for selecting the region of interest for further processing.
Secondly, a classifier is used for distinguishing between the
diseased and non-diseased parts with the selected features as
input. Segmentation can be done by simply considering the
contrast variations in the image, but this will not work well with
images of varying density.This paper tries to find out the best
segmentation method for the image. The classification step is Figure 1. : normal and benign cells
done using Bayesian Classifier
Index terms -. Computed Tomography, Small Cell Lung
Cancer, Non-Smal Cell Lung Cancer,Fuzzy C-Means
There are two main types of lung cancer, small cell (SCLC)
I. INTRODUCTION and non-small lung cancer (NSCLC).About 85% to 90% of
Lung Cancer has the highest mortality rate amongst lung cancers are non-small cell lung cancer, and only 10-15%
all other types of Cancers. Survival from lung cancer is is small cell lung cancer. Following are the subtypes of Non-
directly related to its growth at its detection time. The earlier small cell Lung cancer:
the detection is, the higher the chances of successful treatment • Adenocarcinoma
are. The detection and localization of lung cancer in the micro • Squamous cell carcinoma
invasive stages improves the chances of survival of a patient. • Large cell carcinoma
However, the detection and localization of lung cancer in the
micro invasive stages is very difficult. Small cell Lung cancer spreads faster than non-small cell
Cells are the building blocks of tissues.They are always Lung cancer. In “limited stage”, cancer is found only in the
growing and replace the old ones.There are two types of chest whereas in the ”extensive stage” it spreads outside the
tumors.One is benign and the othermalignant. Cells in benign chest region.
cancer do not spread to other parts of thebody while the
cancerous cells keep on dividing and spread into other parts of
Selection of best imaging modality is the key to detection of
the body in malignant cancer. The spreading of tumor from
the corresponding disease[1]. X-ray, MRI, PET scan and CT
one part of the body to another is called metastasis.
scans are the different imaging modalities currently used.
Among the above, CT scan is used for the three-dimensional
The incidence of lung cancer is strongly correlated with
view of the Lung images[2].
tobacco smoking, with about 90% of the cases arising as a
result of tobacco use. Passive smoking, or the inhalation of
tobacco smoke from other smokers, is also an established risk Segmentation is the process of dividing an image into
factor for the development of lung cancer. Research has shown different regions .The pixels in each group will have similar
that nonsmokers who reside with a smoker have a 24% characteristics. The criteria for grouping the pixels may
increase in risk for developing lung cancer when compared depend on the properties of the pixels. Watershed
with other nonsmokers. segmentation is used for segmenting two touching objects [3].
. For this, topological view of the object is used. Watershed

61
International Journal of Advanced Information Science and Technology (IJAIST) ISSN: 2319:2682
Vol.5, No.5, May 2016 DOI:10.15693/ijaist.2016.v5.i5.61-64

segmentation controls over-segmentation using markers. For

selecting a marker, preprocessing and finding the criteria that
markers should satisfy also have to be considered. The main
advantage of watershed is that, even if there are no strong
edges between the markers, the watershed transform always
detects a contour in the area. The modified adaptive fissure
sweep first pre-processes the CT images to reduce noise.
Wiener filters are used for noise removal instead of a median
filter. The major challenge in this method is the variable shape
and appearance, along with the low contrast and high noise Figure 3. After morphological operation
associated with these images.Channeler Ant Model [4] is
effective whenever complex connected structures are present
in the image. Algorithm consists of selecting the number of
ants. The maximum number of visits for each ant in the voxels
determined primarily. Here selection of number of ants which
is done manually is a challenging task.

Clustering is a method of grouping data objects into different

groups, such that similar data objects belong to the same
cluster and dissimilar data objects to different clusters [5][6].
The algorithm is formulated by modifying the distance
measurement of the standard FCM algorithm to permit the Figure 4. Segmented output
labeling of a pixel to be influenced by other pixels and to
restrain the noise effect during segmentation.FCM fails to Second module is feature extraction, which is used as input to
segment images corrupted by noise, outliers and other imaging the classifier.High-level feature extraction concerns finding
artifact. shapes in computer images. To be able to recognize faces
automatically, for example, one approach is to extract the
Region Growing is an approach in which neighboring pixels component features. Intensity based features, shape, texture
are examined and added to a region class as long as no edges and wavelet are the features extracted here.One of the
are found between them [7]. Choosing region membership is techniques used to improve classification performance is the
more difficult than applying edge detectors. It cannot search inclusion of clustering into the classification process.
objects that span different disconnected regions
In feature extraction, invariance properties are used so that the
extraction process does not vary according to the chosen (or
specified) conditions. That is, techniques should find shapes
reliably and robustly, whatever the value of any parameter that
can control the appearance of a shape.Otsu thresholding is
used for segmenting the CT Lung image.Segmentation is used
to extract features from the selected region of interest.Features
like shape, size, mean, standard deviation and other statistical
parameters are extracted from the segmented region for further
investigation.

Classifier is used for categorizing an unknown pixel into its

Figure 2. input image
corresponding group. Bayesian Classifier is used for
outputting the type of the input image. Classification based on
In morphological operation, segmentation of the lung region is Random forest clustering consists of training and testing set
based on thresholding method [8]. The threshold is determined [9]. Here, the entire data is grouped together and then the
by analyzing the 2D region histogram, which shows distinct dataset is grouped into nodules and non-nodules. It compared
groups of pixels belonging to the thorax and background air. different ratio of training and testing set.

In neuro-fuzzy approach prior knowledge about the training

data set can be encoded into the parameters of the neuro-fuzzy
classifier. Moreover, the parameters obtained after the learning
process can be easily transformed into structured knowledge in
the form of fuzzy if-then rule [5].

62
International Journal of Advanced Information Science and Technology (IJAIST) ISSN: 2319:2682
Vol.5, No.5, May 2016 DOI:10.15693/ijaist.2016.v5.i5.61-64

One of the important problems in fuzzy clustering is how to slicescontaining cancerous nodule is correctly classified as
design membership functions. In this method, based on the cancerous:
input and output of the fuzzy system, neural network is TP
Sensitivity =
TP+FN
trained. The most common neural network model is the
Specificity
multilayer perceptron (MLP). This type of neural network is
It measures the proportion of negatives which are
known as a supervised network because it requires a desired
correctlyidentified. The percentage of segmented slices
output in order to learn. The goal of this type of network [10]
without can-cerous nodule is correctly identified as non
is to create a model that correctly maps the input to the output
cancerous:
using historical data so that the model can then be used to TN
produce the output when the desired output is unknown. Specificity =
TN+FP
Accuracy
Accuracy is a statistical measure of how well a classifier cor-
II.MATERIALS AND METHODS rectly identifies or excludes a condition. The accuracy is
theproportion of true results (both true positive and true
negative)in the population.
TP+TN
A. DATA Accuracy =
TP+TN+FP+FN
The role of a classifier is to correctly identify the group of
an unknown pixel.Bayesian classifier is based on Bayes
Metric Observed value
theorem. The Bayesian Classification is a type of statistical
method for classification. Itpredicts the class or group the Sensitivity 89.4%
given sample belongs to.For this, Bayesian classifier makes Specificity 94.1%
use of the class membership probabilities, such as the Accuracy 91.6%
probability that a given sample belongs to a particular class. Table1: Classification results using Bayesian Classifier for
Bayes Rule is stated as follows: “Given a problem instance to cancerous and noncancerous images
be classified, represented by a vector x=(x1….xn)
representing some n features (independent variables), it
assigns to this instance probabilities VI. CONCLUSION
Features are used as input for the classifier. Bayesian classifier
P = (Ck|x1……xn) for each of the K possible outcomes or is a promising method for correct classification.For this, Lung
classes.” region is extracted from the original CT image. From the lung
The problem with the above formulation is that if the number region, the ROIs were obtained. The nodules are evaluated
of features n is large or if a feature can take on a large number based on the features such as mean, standard deviation,
of values, then basing such a model on probability tables is skewness, kurtosis, fifth and sixth central moment ,then
infeasible. Therefore reformulation of the model is good to subjected to classification to classify the input image. The
make it more tractable. Using Bayes theorem[12], the project may be expanded by adding additional features like
conditional probability can be decomposed as HOG. The existing classifiers can be compared and the better
one may be used for improving the efficiency of the system.
p Ck p(x|Ck )
P Ck x = (1)
p(x)

REFERENCES
the above equation can be written as
[1]. Awais Mansoor , Segmentation and image analysis of
prior x likelihood abnormal lungs at ct: Current approaches,challenges, and
posterior = future trends, Radiographics ,2015.
evidence
[2]. Elisabeth Brambilla H.,Konrad Müller-HermelinkCurtis C,
III.EXPERIMENTAL RESULTS Harris William D,Travis,Pathology and genetics of tumours
of the lung, pleura, thymus and heart, WHO Publications
Center, 2004.
The training set for the Bayesian classifier consists of70% of [3]. Lung lobe segmentation by anatomy-guided 3d watershed
the total images and the testing set is 30% of the total images. transform, Medical Imaging 2003: Image Processing, Vol.
The sensitivity, specificity andaccuracy are calculated as 4, No. 2, pp. 1482-1490,2003.
follows: [4]. Piergiorgio Cerello and Sorin Christian Cheran, 3-d object
segmentation using ant colonies,IEEE Nuclear science
Sensitivity Symbosium Conference, ISSN :1082-3654, 2008.
[5]. S.Sivakumar and Dr.C.Chandrasekar,Lung nodule
It measures the proportion of actual positives which are detection using fuzzy clustering and Support vector
correctly identified. That is the percentage of segmented

63
International Journal of Advanced Information Science and Technology (IJAIST) ISSN: 2319:2682
Vol.5, No.5, May 2016 DOI:10.15693/ijaist.2016.v5.i5.61-64

machines,International Journal of Engineering and

Technology (IJET) 5 (1), 179-185, 2013.
[6]. R.E. Wood, R.C.Gonzalez.,Digital image processing, ISBN
number 9780131687288,3rd edition, Prentice-Hall,2008.
[7]. Atiyeh Hashemi., Mass detection in lung ct images using
region growing segmentation and decision making based
on fuzzy inference system and artificial neural network,
I.J. Image, Graphics and Signal Processing, 2013, 6, 16-
24.
[8]. Jayashree,P Sudha.V, Lung nodule detection in ct images
using thresholding and morphological operation,
International Journal of Emerging Science and Engineering
(IJESE) ISSN: 2319–6378, Volume-1, Issue-2, December
2012.
[9]. S.L.A. Lee., A random forest for lung nodule identification.,
Computerized Medical Imaging and Graphics 34 (2010)
535– 542,2010.

[10]. M.G. Penedo, Computer-aided diagnosis a neural-network-

based approach to lung nodule detection,IEEE
Transactions Medical Imaging . 1998 Dec;17(6):872-80
[11]. M. Madheswaran and D. Anto Sahaya Dhas, Classification
of brain MRI images using support vector machine with
various Kernels, Biomedical Research 2015; 26 (3): 505-
513
[12]. Naïve Bayesian Classifier-English Wikipedia.
https://en.wikipedia.org/wiki/Naive_Bayes_classifier.

Authors Profile
Hitha Rocky is currently doing her
master’s degree in Technology,
specializing in Computer Science and
Engineering at Adi Shankara Institute of
Engineering and Technology, Kalady. Her
areas of interest include image processing,
Neuro-Fuzzy

Dr.Jereesh A.S is currently working at

Cochin university of Science and
Technology as Assistant Professor in
Computer Science Department.
Received Phd from NIT.His area of
interest is Image Processing.

64