You are on page 1of 6

2017 Fourth International Conference on Image Information Processing (ICIIP)

$Q2SWLPL]HG/XQJ&DQFHU&ODVVLILFDWLRQ
6\VWHPIRU&RPSXWHG7RPRJUDSK\,PDJHV
Sheenam Rattan6XPDQGHHS.DXU
1LVKX.DQVDO-DVSUHHW.DXU
Dept. of ECE GNDEC/udhiana-141006 India
VKHHQDPUDWWDQ#JPDLOFRPVXPDQVDUDRQ#JPDLOFRPQLVKXNDQVDO#JPDLOFRP
MDVSUHHWNDXU#\DKRRLQ

Abstract:- Amongst diverse cancers, lung away through lung cancer. Evidences of cancer
cancer is measured to be the foremost reason of includes breath conciseness, out of breath, upper
cancer demise with utmost demise pace. Nodules body throbbing which is no more getting healthier,
lying on lungs have distinct structures, they choking accompanying red liquid, deglutition
could be either circle or coil shaped which under complexity, thrashing of weight along with appetite
various circumstances composes the recognition [16]. There are considerable facts representing that
complex. In this work a system has been mortality pace can be reduced by before time
urbanized for detection of lung cancer in its recognition about lung cancer. Current assessment
early stages and classification between given by recent data presented via health
malignant and benign tumors via images from association reveals that due to lung cancer about
Computerized Tomography (CT) scanner. Lung 7.6 million demises all around the world. In
cancer detection process has four steps which addition, fatality coming out of tumor is headed for
include pre-processing phase, segmentation, growing, and turns out to be around 17 million
feature extraction and lung cancer cell globally in 2030. A variety of treatments for curing
classification. BAT Algorithm is applied to this cancer includes radiotherapy, chemotherapy as
provide considerable optimization results which well as surgery.For lung cancer, 5 year endurance
improve the performance of system. The speed meant for these patients has been fallen down
classification between malignant nodules and to 14%.Conversely, during another types of cancer,
benign has been done through Artificial Neural endurance pace may possibly rise and reaches
Network Ensemble to provide results of higher about 49% of the condition that it is recognized
accuracy. The overall accuracy, sensitivity and before time.
specificity of 98.5%, 100% and 91% respectively To identify lung cancer, with utmost persistence
is acquired in the system. Computed Tomography be used. Medicinal
leftovers as well as nodules with different length
can be contentedly examined with Computed
Keywords:- Lung cancer, CT Images, Tomography [3]. Lumps present above the lobe of
Watershed Transformation, BAT Algorithm, lung are categorized as malignant else benign. In
Graylevel co-occurrence matrix (GLCM), few situations lumps that are firm as well as
Artificial Neural Network Ensemble (ANNE). uncommon are termed as malignant and are
evaluated like benign through diagnosis [4]. So
1. Introduction diagnosis before time is vital so as to increase
These days, lung cancer is the foremost part of speed of medical aid.
lethal types of cancer. Furthermore it is Computer Aided Diagnosis technique planned in
complicated to perceive in its early stages for the order to various surgical applications gives distinct
reason that signs emerges in its higher stages which advantages as thriving recognition about lumps.
causes the demise rate to be maximum when With these techniques medication procedure should
compared to all other kinds of cancer. In be started before time as well as help in directing
comparison with diverse cancer forms like thyroid, procedure for surgeons. Such mechanized system
pancreatic and bladder, the most population passes would allocate objective and unbiased estimation,

978-1-5090-6734-3/17/$31.00 ©2017 IEEE 15


2017 Fourth International Conference on Image Information Processing (ICIIP)

as compared to human appraisal which may be positioning, but while considering the current
corrupted with errors. In the long run, this system problem the limited deviation in an image is
will be helpful for managing large image database directed by bats. To obtain the results practical
as well as relieving the pathologist from outcomes have been featured out.
monotonous routine task.
2. Literature Survey 3. Present Work
The researcher in the paper [14] urbanized different In this paper, BAT algorithm has been used to
technique about automated recognition of nodules provide optimal solutions and further Watershed
of benign. Genetic algorithm template matching Transformation has been used for nodule
(GATM) which is a new template matching method segmentation. GLCM technique has been applied
is used to identify nodules present inside the area of for the feature extraction. ANNE has been engaged
lung Detection of lung nodules performed in this for classification. ANNE is preferred as with the
method provides accuracy of 72% approximately. increase in complexity of the system it performs
The researcher in the paper [8] projected a model better. ANNE has the capability of making the
based on ANNE. The ensemble made comprises of system easier to interpret or modify the model if
two steps in which first level of ensemble is used to required.
find out if cancer is present or not. The second The planned system consists of four main phases:
level ensemble further deals with the cells that are 1. Pre-processing
termed as cancerous. By using neural network 2. Region of Interest (ROI) to be Segmented
ensemble, results in providing higher rate of overall 3. Extraction of Features
classification of disease as well as lowers the rate 4. Categorization of Malignant or Benign nodules
of false negatives. The researcher in the paper [2] Fig.1 shows original image of lung which is
planned a CAD scheme so as to identify malignant diagnosed to detect if image is normal or abnormal.
lung nodules by using CT images. To get better
images morphological operators are used in order
to increase the partition of regions with
dissimilarity as well as involves filling of gaps
present in the lumps. After performing this scheme
nodules containing malignant is detected and false
positive identification rate is also decreased. The
identification of nodules provides an accuracy of
65.4% approximately. The researcher in the paper
[13] explained categorization scheme using ANN
on images. Morphological operators are applied for
segmenting the CT images. Otsu method is used to
calculate the threshold value. Early diagnosis of
nodules of lung is also not focused. In this study
suggestion regarding diagnosis of nodules small in
size is not given. The researcher in the paper [4] Fig.1 Original image
projected a method which provides before time
detection of disease, identifying different Each step is discussed below:
treatments required at different stages as well as 3.1 Pre-processing:-The processing of an image
decreases the mortality rate. In order to classify begins with enhancing the image, the main idea
cancer as malignant or benign in the beginning behind enhancing an image involves improvement
stage an artificial neural network has been in the image to make it more illustratable and get
proposed. To extract the features of nodules GLCM better data observations included in an image and
scheme has been applied. For categorization step also provides improved input to other mechanized
artificial neural network is applied. It involves processing of an image scheme. Quality of an
processing of an image which is applied to improve image is enhanced as well as noise is removed by
the quality of an image as well as removes noise processing of an image. For removal of noise and
present in an image. The nodules of lung provide enhancing the image Morphological operators are
an accuracy of 93.3%. The researcher in the paper used. Fig.2 shows the first step in which image is
[7] projected detection of edges by using BAT pre-processed and enhanced image is obtained. To
algorithm. BAT algorithm is considered as nature allocate pixel values having stability histogram
inspired algorithm pulse rate emission and equalization is applied. Fig.3 displays the
echolocation is used by bats to locate their con and histogram of the image that is pre-processed.
find out the hurdles in their path. Same set of data
is used so as to classify their prey in situation
where there is complete black out. The main
advantage of BAT algorithm is aimed at bats

16
2017 Fourth International Conference on Image Information Processing (ICIIP)

r€[0,1] can be adjusted relying on closeness of the


prey.
3. Even though loudness could be varied in
numerous manner, but assumptions are made that
loudness varied from large value A0 to the smallest
value Amin. Initializing the population of bat is
conducted arbitrary, in BAT algorithm. New
solutions are generated by operating bats in
accordance with equations (1.1), (1.2),(1.3):
x fi = fmin+ (fmax-fmin)β (1.1)
x vit = vit-1+ (xit -x*)fi (1.2)
x xit = xit-1+ vit (1.3)
Fig.2 Shows Enhanced image Where,β € [0, 1] refers to a vector that is taken
randomly from standardized allocation. Now the
present global best solution is denoted by x* which is
the result that comes out after comparing results of
various solutions provided by bats. During the
implementation part fmin=0 and fmax=2 is used, which
is dependent on the area range of that particular
problem. Every time bats are assigned with a random
frequency that can be drawn from [fmin, fmax]. For
local search which results in modification in best
solution is directed by random walks. To modify the
current best solution random walks are used for local
searching. While loudness parameter is denoted by
At. The search termed as local search starts with the
closeness that depends on the pulse rate emission ri.
Fig.4 (a) shows open area of an image and (b)
shows segmentation of nodules of lung using
Fig.3 Shows Histogram of pre-processed image Watershed Transform scheme, by using this
scheme the region containing cancer is extracted
3.2 Segmentation of nodules of lung:-Segmenting out from the image. Fig.5 shows the optimized
an image aims to partition computerized image into results of the function that are segmented.
significant regions considering some factors like
intensity, colour or texture. Segmenting an image is
a crucial process and is used in large number of
medical treatments. For segmentation of image
Watershed transform has been used. To find the best
solution optimization technique i.e. BAT algorithm
has been employed. It is an algorithm inspired by
nature. Wavelength (λ) that is being used by bat
ranges from 0.7 to17 mm and the frequency (f)
ranges from 20-500 kHz. Frequency (f) can be varied
by fixing the wavelength, f and λ are related to each
other because of the fact λf is constant. Rate of pulse
can be directly decided from 0 to1 a given range,
where 0 stands that no emission is taking place and 1
stands for that emission is maximum. While
implementing the BAT algorithm the rules being
generalized are:-
1. Echolocation is used by bats to detect the distance,
bats in some magical way have the ability to
differentiate among food or pray as well as
surrounding obstacles.
2. In order to search food bat flies with random
velocity (vi,) at a given position (xi) having frequency
which is fixed fmin, with varying wavelength λ as well
as loudness A0. Wavelength of the pulses that are
emitted can be adjusted automatically and pulse rate Fig.4 (a) Areaopen Image (b) Watershed Transform

17
2017 Fourth International Conference on Image Information Processing (ICIIP)

Adoption of neural network ensemble not only


provides higher accuracy also lowers the
identification of false positive which means
lowering the judgment of cells of cancer as normal
cells, that helps to save lives of many patients of
cancer by providing diagnosis in early stages. Fig.6
(a) & (b) shows the output that is obtained from
first level of ensemble i.e. if cancer is present in the
image or not, Fig.7 (a) & (b) shows final output
image depicting the type of cancer. NNE is an
approach formed to solve complicated problems.

Fig.5 Optimized Function

Fig.6 (a) Output of first stage of ensemble having tumor (b)


3.3 Extraction of features:-While constructing Output of Output of first stage of ensemble having no tumor
any classification pattern extraction of features is
considered as an essential step. Feature extraction
provides extraction of information that is
significant. Feature extraction is defined as process
in which input information is transformed into
relevant set of data. It is considered as a process in
which most significant data is recovered from set of
unprocessed data. In this method gray level co- Fig.7 (a) Malignant Cancer (b) Benign Cancer
occurrence matrix (GLCM) based scheme is used.
The main advantage of using gray level co- 4. Experimental Results and Sensitivity
occurrence matrix over other schemes is that it
involves reduction in compression time for Analysis
converting a RGB image to Gray level image. 4.1 Complete Diagnosis Rate of the system
Energy, correlation, homogeneity and contrast are A total of 70 images are used, out of which 10
the parameters extracted using GLCM. images are normal and 60 images are diseased. Out
of the diseased images, 35 images are malignant
3.4 Classifying lung nodules using artificial and 25 are benign. 10 images not containing cancer
neural network ensemble:- In this approach a in it are correctly recognized. All 35 images related
single problem is broken into sub groups further to malignant cancer are correctly recognized. On
solved by solvers in groups. Neural network the other hand 24 images related to benign cancer
ensemble produces better and more generalized are correctly recognized but only 1 is misclassified
results as compared to single artificial neural as malignant. Table 1 depicts total number of images
network. It is considered as a model of learning in taken, their correct detection and error rate. Fig.8
which a lot of networks of neural are joined so as to represents the graph depicting the total number of
find solution of a particular problem. Selecting normal, malignant and benign images used their
weights in an optimized problem having a large correct detection (%) and error rate.
number of local minima is difficult but ensembles
Table 1 Complete Diagnosis Rate of the system
provide desirable results. This difference in weights
leads to several other ways in generalizing the Criteria for Normal Malignant Benig
patterns involved in training. Two-level Assessment n
architecture of ensemble has been built up. In the Number of 10 35 25
very first level neural network ensemble is used to images
deduce if the cancer cell is ordinary containing high Correct 100 100 96
assurance, 2 outputs have been resulted from Detection (%)
respective network which includes whether there is Error Rate 0 0 4
tumor present or whether there is no tumor. The
cells which are deduced of having cancer in it by
the first level is then further handled by the second
level of ensemble , similarly two outputs are
resulted from respective networks which includes
type of cancer i.e. malignant or benign cancer.

18
2017 Fourth International Conference on Image Information Processing (ICIIP)

120 Table 3 Criteria for Assessment


100 Number of
Criteria for Assessment Results (%)
80 images
Accuracy 98.5
60
Correct Sensitivity 100
40 Detection Specificity 91
20 (%)
0 Error Rate
102
100
98

Scale(%)
96
94
Fig.8 Diagnosis Rate of the system 92
90
4.2 Analysis of overall performance of the 88 Results (%)
system: 86
To measure the overall performance of the system,
three parameters are considered which includes
accuracy, sensitivity and specificity. Accuracy is
used to measure percentage of correctly classified
images, Sensitivity is used to measure true positive Parameters
ratio of classified images, Specificity is used to
measure true negative ratio of the classified images.
Table 2 represents the confusion matrix in which Fig.9 Overall performance of the system
TP, FP, TN and FN are represented. Here TP, TN,
FP, FN represent the number of true positive, true The present work proposes a system that
negative, false positive and false negative samples successfully distinguishes the lung nodules as
respectively. TP and TN indicate when the benign or malignant. The present work provides
classifier is receiving things right, while FP and FN early detection of lung cancer with higher accuracy
indicate when the classifier is receiving things rates. The most important contribution of this work
wrong. P is the number of a positive set of elements is that it allows detection of very small sized
and N is the number of a negative set of elements. nodules, so that the cancer can be cured easily
In present work diseased images are termed as reducing death rate.
positive set of elements and normal images are
termed as negative set of elements. Table 3 5. Conclusion
represents different parameters of comparison. The CT images have been used in the present
Fig.9 represents the graph depicting overall system to successfully distinguish the lung nodules
performance of the system. as benign or malignant. It has been concluded that
using the designed system the mortality rate can be
Table 2 Confusion Matrix reduced as cancer present can be detected in its
early stages. Pre-processing, lung nodule
Predicted
Yes No
Total segmentation, extraction of features as well as their
Yes TP FN T classification are the various steps that are involved
Actual No FP TN P in the system. Various features are extracted from
the image using GLCM. BAT algorithm is used as
Total T’ P’
T+P or an optimization technique for segmentation
T’+P’ process. At classification stage, ANNE technique is
used to achieve better accuracy. By using the
TP=59, TN=10, FP=1, FN=0 technique presented many other types of cancer can
also be detected. With this system, the obtained
TP + TN
ccuracy = = 98.5% accuracy is 98.5%, sensitivity is 100% and
TP + FN + TN + FP specificity is 91% respectively, which offers
significant improvement over the existing systems.
TP
Sensitivity = = 100%
TP + FN
TN
Specificity = = 91%
TN + FP

19
2017 Fourth International Conference on Image Information Processing (ICIIP)

8. Zhou, Z., Jiang, Y., Yang, Y. and Chen, S.


6. Future Scope (2002), “Lung Cancer Cell Identification
Based on Artificial Neural Network
x Enhancing radiologic techniques may possibly
Ensembles”, Artificial Intelligence Medicine,
provide diagnostic information more
vol.24, pp. 25-36.
accurately.
9. Fister, I., Izotak, Yang, X., Fong, S., Zhuang,
x Predictable studies in complement with
Y. (2014), “BAT Algorithm: Recent
molecular diagnosis might possibly have more
Advances”, Computational Intelligence
impact on supervision of patient.
Informatics IEEE 15th International
x Combining various features that already exist Symposium, pp. 163-167.
so that more accuracy can be obtained while 10. Gupta, B. and Tiwari, S. (2014), “Lung
classifying medicinal data. Cancer Detection using Curvelet Transform
x So as to decrease the mortality rate and and Neural Network”, International Journal
increase the life span of patients ANT Computing Applications, vol. 86, pp. 15-17.
COLONY along with or in combination of 11. Hussain, A., Ansari, M., Gawas, S. and
NEURAL NETWORK can be done. Chowdhury, N. (2015), “Lung Cancer
x In the current scenario processing of an image Detection using Artificial Neural Network &
is done on gray scale images further Fuzzy Clustering”, International Journal
improvement can be done directly on images Advanced Research Computer
those are colored. Communications Engineering, vol. 4, pp. 136-
142.
REFERENCES 12. Kanazawa, K., Kawata, Y., Niki, N., Satoh, H.,
1. Alihodzic, A. and Tuba, M. (2013), “BAT Ohmatsu, H., Kakinuma, R., Kaneko, M.,
Algorithm (BA) for Image Thresholding”, Moriyama, N. and Eguchi, K. (1998),
Recent Researches in Telecommunications, “Computer Aided Diagnosis for Pulmonary
Informatics, Electronics and Signal Processing, Nodules Based on Helical CT Images”,
pp. 17-19. Computing Medical Imaging Graph, vol. 22,
2. Biradar, V. and Patil, U. (2013), “Computer pp. 157-167.
Aided Detection System for Automatic 13. Kuruvilla, J. and Gunavathi, K. (2014), “Lung
Pulmonary Nodule Detection in Lungs in CT Cancer Classification using Neural Networks
Scans”, International Journal Engineering for CT Images”, Computing Methods
Science, vol. 2, pp. 18-21. Programs Biomedical, Elsevier, vol.113, pp.
3. Choi, W. and Choi, T. (2013), “Automated 202–209.
Pulmonary Nodule Detection System in 14. Lee, Y., Hara, T., Fujita, H. and Ishigaki, T.
Computed Tomography Images: a (2001), “Automated Detection of Pulmonary
Hierarchical Block Classification Approach”, Nodules in Helical CT Images Based on an
International Journal Biomedical Imaging, vol. Improved Template-Matching Technique”,
15, pp. 507- 523. IEEE Transactions medical imaging, vol. 20,
4. Dandıl, E., Ozkan, M., Eksi, Y. and Canan, A. pp. 595-604.
(2015), “Artificial Neural Network-Based 15. Miah, B. and Yousuf, M. (2015), “Detection of
Classification System for Lung Nodules on Lung Cancer from CT Image Using Image
Computed Tomography Scans”, 6th Processing and Neural Network”, 2nd
International Conference Soft Computing International Conference on Electrical
Pattern Recognition, Tunis, Tunisia, pp. 382- Engineering Information Communication
386. Technical, Dhaka, Bangladesh, pp. 21-23.
5. Dhar, S., Alam, S., Santra, M., Saha, P. and 16. Phyllis and Li, F. (1999), “Annual Report to
Thakur, S. (2017), “A Novel Method for Edge the Nation on the Status of Cancer, with a
Detection in a Gray Image Based on Human Special Section on Lung Cancer and Tobacco
Psychovisual Phenomenon and BAT Smoking”, Journal National Cancer Institute,
Algorithm”, Computer, Communication vol. 91, pp. 675-690.
Electrical Technology, pp. 960-978. 17. Rattan, S., Kaur, J. and Kansal, N. (2017),
6. Dolejsi, M. and Itoh, S. (2007), “Detection of “Artificial Neural Network-Based System for
Pulmonary Nodules from CT Scans”, MSc Lung Nodules for Early Detection of Lung
Thesis, Department Electrical Engineering, Cancer”, International Conference on Soft
Czech Technology University, Prague, Europe. Computing Applications in Wireless
7. Dwivedi, R., Sethi, H., Rohilla, M., Garg, V. Communication, Ludhiana, India, pp. 1-9.
and Nagpal, Y. (2016), “Edge Detection using
BAT Algorithm”, 3rd International Conference
Computing Sustainable Global Development,
New Delhi, India, pp. 2605-2608.

20

You might also like