IJSTE - International Journal of Science Technology & Engineering | Volume 3 | Issue 09 | March 2017

ISSN (online): 2349-784X

Classification of Lung Tumor from CT Images

using Computer Aided Diagnosis Scheme
Amritha Vijayan
PG Student
Department of Electronics & Communication Engineering
Marian Engineering College, Trivandrum, India
Minnu Jayan. C
Assistant Professor
Department of Electronics & Communication Engineering
Marian Engineering College, Trivandrum, India

Abstract
Lung cancer is the second most prevalent type of cancer. Early detection and classification of lung tumors helps in better survival.
The paper presents a computer aided diagnosis (CAD) system for chest computed tomography (CT) images that focus on a new
algorithm to classify lung tumor into benign or malignant using advanced neural network based classifier. Statistical textural
features are extracted which aids in better classification. The accuracy of the algorithm is also computed. Methods: Input image is
pre-processed by median filtering and segmented to extract the portion of lungs. This is followed by identifying the tumor objects
alone from the region of interest of the lungs. Statistical textural features are extracted only from the tumor objects. Classification
of tumor into benign and malignant tumor is done using feed forward back propagation neural network classifier. The result shows
proposed algorithm gives classification accuracy above 99% which suggests that the developed CAD system has greater potential
for automatic classification of lung tumor.
Keywords: Artificial Neural Network (ANN), Benign Tumor, Computer Aided Diagnosis (CAD), Computerized
Tomographic Images (CT), Levenberg Marquardt Training Algorithm, Malignant Tumor
________________________________________________________________________________________________________

I. INTRODUCTION

Lung cancer is the second most prevalent type of cancer in both men and women [1]. It was estimated by cancer society that 14%
of all new cancers are lung cancer. Approximately one out of four cancer deaths are due to lung cancer. It results in higher mortality
rate in recent years with rate higher rate than breast cancer, colon cancer and many other type of cancers. It accounts to 1.59 million
deaths per year. Most people who develop lung cancer do not develop clinical symptoms until it become advanced as many people
mistake them for other problems such as infections [2]. Being a relatively large body human organ that may involve many other
chronic diseases, when tumor is found it had already resulted in 95% loss of pulmonary function. Early detection of lung cancer
helps in early treatment and increases survival rate up to 80%. The common symptoms of lung cancer are cough, dyspnea,
hemoptysis and systemic symptoms such as weight loss and anorexia [3]. It is mainly caused due to smoking, exposure to certain
type of chemicals, family history. Abnormal rates of cell division of lung tissues results in lung tumor. There are mainly two type
of lung tumor based on whether the cells invade other parts of body [4]. It is benign tumor if it does not invade its surrounding
tissue and malignant if it spreads to other parts of the body. If the cells of the malignant tumor reaches other parts of the body
through blood or lymphatics, that area will also be affected (metastasis). Hence it is important to find the malignancy of tumor
before it spreads, resulting in multiple cancer and death. Based on histology lung cancer is classified into small cell and non-small
cell lung cancer. The main imaging modalities to image lungs are chest radiography, computed tomography (CT), Magnetic
resonance Imaging (MRI scan) [5]. Screening using CT is on the rise due to low cost, robustness and quality. A part of detected
tumor can be cancerous. But for screening purposes CT with high radiation is harmful and hence recently developed low dose CT
(LDCT) with less radiation. For detecting extra-encephalic malignant tumors and unexpected lesions, CT images shows high
performance up to 30% [6]. It make use of the image processing techniques to assist radiologists to interpret the medical images
and helps in diagnosis with high accuracy by avoiding chances of human error [7,8]. Based on the quantitative tumor related
features even small lesions can be detected using CAD scheme. This paper proposes computer aided diagnosis system for
classification of lung tumor with high accuracy using neural network based classifier.

II. RELATED WORK

To state the problem of automatic detection and classification of lung tumors, researchers had proposed several methods. The aim
of all these researchers is to develop a system which can predict, detect and classify tumor in its early stage with high accuracy.
However, all of their works involved these main steps. The steps are pre-processing, segmentation, feature extraction and
classification. Some of the major research works are summarized below.
Tiwari [9] provided a review of the various image processing techniques and pattern recognition based approaches by the
previous researchers in the prediction of lung cancer. A comparison of different classification techniques and their accuracies
proposed by different researchers were also presented.

All rights reserved by www.ijste.org 589

 Classification of Lung Tumor from CT Images using Computer Aided Diagnosis Scheme
(IJSTE/ Volume 3 / Issue 09 / 118)

The CAD system developed by Jacobs et al [10] focus on the detection of sub-solid pulmonary nodules by using 128 image
features. Besides using the existing features such as intensity, shape and texture features, context based features were also extracted.
The combination of these set of features significantly improved the classifier performance. The developed system finally reaches
a sensitivity of 80%.
Samuel H Hawkins et al [11] developed a CAD scheme using selected features from CT images for the prediction of lung cancer
using different classifiers. Rule based, decision tree, Naive Bayes and support vector machine classifiers are compared for their
accuracies in cancer prediction. Among them, decision tree classifier is found with best predictive accuracy.
Computer aided classification method by Frahani et al [12] is based on an ensemble of three classifiers. The classifiers includes
multilayer perceptron, K-Nearest neighbor and support vector machines. Shape related features such as roundness, compactness,
circularity, and eccentricity were extracted from the segmented part of lungs. In order to combine the result of these classifiers,
majority voting method was used. The ensemble system showed high performance in the detection of pulmonary nodules from CT
images.
Sheeraz Akram et al [13] had proposed an automatic pulmonary nodule detection system with the aid of artificial neural
networks. The CAD system used a standard Lung Image Consortium Datasbase (LIDC). The lung volume is segmented out and
the candidate nodules are extracted from the lungs based on the rules which are formed using the characteristics of nodules. The
neural network is trained using geometric and intensity based statistical features. The developed CAD system has highest
performance when compared to other classifier based CAD systems.
F. Paulin et al [14] studied the classification of breast cancer using different feedforward neural networks. The network
performance is computed using various back propagation training algorithms. The highest accuracy of 99.28% is obtained when
using the Levenberg Marquardt (LM) training algorithm. Nawaz Khan et al [15] provides a case study for the performance
evaluation of LM algorithm in the error reduction of diabetes condition classification. A dynamically constructed neural network
is trained and tested by different back propagation algorithms and a comparison of their output is done in terms of accuracy,
sensitivity and specificity. From the implementation, experimentation and performance evaluation, it is found that LM algorithm
outperforms all other training algorithms.

III. PROPOSED METHOD

The proposed CAD system is developed to classify the lung tumor into benign or malignant. The aim of the system is to minimize
the computational time and to obtain the output with high algorithm accuracy. The CAD system is implemented in MATLAB
using image processing and highly advanced neural network classifier. The classifier is validated using leave-one-out validation
technique. The proposed CAD system consist of various stages which includes the image acquisition, pre-processing,
segmentation, morphological filtering, region of interest selection, identifying tumor regions from the region of interest, feature
extraction and classification. The block diagram of proposed system is shown in Fig.1

Image Acquisition
The images are obtained from the public access research lung image database that was available to be used for performance
evaluation of different CAD systems. The database consist of 35 low dose CT scans which were obtained in a single breath hold.

Pre-Processing
This stage aims to enhance the quality of image so as to increase the precision and accuracy in the later stages of image processing.
Pre-processing removes the noise occurred during the image acquisition process. Using median filter, the unwanted noise
components are removed while preserving the edges.

Segmentation
It is an important step in the image processing of medical images. It is done to extract only the portion of lungs by separating the
objects and boundaries. It is done by initially converting the pre-processed grayscale image into binary image by thresholding. The
binary image is then filled to obtain the boundary region. The filled image is subtracted from the binary image to extract only the
portions of lungs. The complement of the extracted portion of lungs is then obtained.

Morphological Filtering
The contrast and quality of the extracted portions of lungs is enhanced by morphological filtering [16]. This is done by erosion
followed by dilation. Erosion shrinks and removes objects of certain size and shape described by the structuring element. While
dilation grows objects with certain size and shape. If the pixels have low intensity neighborhood, then their intensity values are
decreased by erosion. While dilation does the opposite by increasing the intensity values of pixels with high intensity
neighborhoods.

All rights reserved by www.ijste.org 590

 Classification of Lung Tumor from CT Images using Computer Aided Diagnosis Scheme
(IJSTE/ Volume 3 / Issue 09 / 118)

Fig. 1: Block diagram of proposed algorithm

Region of Interest (ROI)

Region of interest is the area of the lungs where further focus of image analysis is required. It is graphically selected from the
window displaying the image as a polygonal area of interest.

Identifying the tumor regions from the region of interest

The tumor regions from the region of interest are identified for further feature extraction and classification. This is done by
extraction of connected objects from the displayed region of interest by making use of the fact that the tumor regions tend to be
brighter than rest of the regions in the selected part of left or right lung.

Feature Extraction
Among the different type of image features available, histogram based statistical textural features are extracted. Textural features
are selected as their orientation is different for different classes. The first order statistical features are extracted in this study [17].
The statistical features are extracted from the tumor regions alone. These features are used for training. The various first order
statistical features used in the study are
Mean
It is the measure of central tendency of probability distribution. The mean of all the pixel values of the tumor region is calculated.
The mean of the tumor region is calculated by using (1)
1
= M N P(i, j) (1)
MN i=1 j=1
Where P(i,j) is the intensity value at the point (i,j) and M*N is the size of the image.
Variance
It is comparatively more incredible feature than mean. It gives the amount of grey level fluctuations from the mean grey level
values. The variance of the tumor region is calculated by using (2).
1
2 = M N (P(i, j) )2 (2)
MN i=1 j=1
Skewness
It is the measure of asymmetry of probability distribution. It measures if there is a wider range of either lighter pixels or darker
pixels. The skewness of the tumor region is calculated by using (3).
1 P(i,j) 3
S= M N
i=1 j=1 [ ] (3)
MN
Kurtosis
It is a measure of peakness of probability distribution. Also it is a measure of the shape descriptor of probability distribution. The
kurtosis of the tumor region is calculated by using (4)
1 P(i,j) 4
K= M N
i=1 j=1 [ ] (4)
MN

Classification of Tumor
Classification of tumor into benign or malignant type is done using feedforward backpropagation neural network. It outperforms
other classifiers in terms of classification accuracy for binary classification [18]. Back propagation network consist of input layer,
hidden layer and output layer. Back propagation is based on error correction rule and the weights are adjusted by back propagating
the associated error. The input database is divided into training, testing and validation set. The classifier is trained using the
statistical textural features of the tumor using the training dataset. Levenberg Marquardt backpropagation algorithm is used for
training the classifier. It is the fastest training algorithm available in the toolbox and helps in obtaining highly optimized output
[19]. Training the classifier is followed by testing the classifier with a new test input. Classification of tumor in the region of
interest of the test input is done by recognizing trained patterns by the network. Validation dataset is used to measure the
generalization capability of network to the new test inputs after training. Validation is done by leave-one-out validation technique.
When the generalization stops improving, the training is stopped. The algorithm accuracy is computed and performance of the
network is plotted with mean square error (mmse) against the number of epochs. The performance plot shows the rapid drop in

All rights reserved by www.ijste.org 591

 Classification of Lung Tumor from CT Images using Computer Aided Diagnosis Scheme
(IJSTE/ Volume 3 / Issue 09 / 118)

mean square error as the network learns. In the performance plot, blue line shows the decreasing error on the training data while
the red line shows the error on test data. The green line shows the decreasing error on validation data as the network generalizes
to new data.

IV. RESULTS
The given image of lungs with tumor is converted initially into a binary image as given by Fig. 4a). The binary image is filtered
using median filter. The filtered image is then filled to obtain the boundary region. The obtained boundary region is subtracted
from the binary image to extract the lungs shown in Fig. 4b). The complement of Fig. 4b) is taken and complemented image is
morphologically filtered and the finally obtained image is as shown in Fig. 4c). After morphological filtering the region of interest
is selected given by Fig 4d). Then tumor regions are identified from the region of interest shown in Fig. 4e). It is followed by
extraction of features from the tumor as shown in Fig. 4f). It is followed by the classification of tumor into benign or malignant as
Eroded
subtracted image
shown in Fig. 4g). The performance plot is shown in Fig. 4h) with training stopped at epoch 4 as generalization stops improving.
Binary Image

Fig. 4(a): Binary image Fig. 4(b): Extracted lungs Fig. 4(c): Morphologically filtered image
Select Part of Left Lung/ Right Lung)

Fig. 4d) Region of interest Fig. 4e) Extracted tumors Fig. 4f) Extracted features

Fig. 4(g): Classification into benign or malignant tumor

All rights reserved by www.ijste.org 592

 Classification of Lung Tumor from CT Images using Computer Aided Diagnosis Scheme
(IJSTE/ Volume 3 / Issue 09 / 118)

0
Best Validation Performance is 0.41644 at epoch 4
10
Train
Validation
Test
Best

Mean Squared Error (mse)

-5
10

-10
10

-15
10

-20
10
0 0.5 1 1.5 2 2.5 3 3.5 4
4 Epochs
Fig. 4(h): Performance plot

V. CONCLUSION

The result shows that the developed CAD system gives classification accuracy above 99% which suggests that the developed CAD
system has greater potential for automatic classification of lung tumor.

REFERENCES
[1] American Cancer Society, Cancer Facts & Figures, 2016. [Online].Available: https://www.cancer.org/research/cancer-facts cancer /cancer-facts-figures
2016.htm.
[2] Lung cancer screening and early detection-American Lung Association 2015. [Online]. Available: http://www.lungforce.org/early-detection.
[3] Kelly M. Latimer, MD, MPH, and Timothy F. Mott, MD Lung Cancer: Diagnosis, Treatment Principles, and Screening, Am Fam Physician., 15;91(4):250-
256, 2015 Feb.
[4] M. Riihimaki A.B, et al; Metastatic sites and survival in lung cancer, Elsevier, Lung Cancer 86 (2014) 7884, July 2014.
[5] Bhavanishankar. K1 and Dr. M.V.Sudhamani, Techniques for detection of solitary pulmonary nodules in human lung and their classifications - a survey,
International Journal on Cybernetics & Informatics (IJCI) Vol. 4, No. 1, February 2015.
[6] David Ladron de Guevara Hernndez, The Role of PET/CT Imaging in Lung Cancer, Journal of Cancer Therapy, 6, 690-70, 2015
[7] K DOI, Current status and future potential of computer-aided diagnosis in medical imaging, The British Journal of Radiology, 78 (2005), S3S19 E 2005,
The. British Institute of Radiology DOI: 10.1259/bjr/82933343.
[8] Ayman el-baz, et al; Computer-Aided Diagnosis Systems for Lung Cancer: Challenges and Methodologies, International Journal of Biomedical Imaging,
Volume 2013 (2013), Article ID 942353.
[9] Arvind Kumar Tiwari, Prediction of lung cancer using image processing techniques: a review, Advanced Computational Intelligence: An International
Journal (ACII), Vol.3, No.1, January 2016.
[10] Colin Jacobs, et al; Automatic detection of subsolid pulmonary nodules in thoracic computed tomographic images, Elsevier, Medical Image Analysis 18
(2014) 374384, December 2013.
[11] Samuel H. Hawkins, et al; Predicting Outcomes of Nonsmall Cell Lung Cancer Using CT Image Features, IEEE Access, December 10, 2014.
[12] Farzad Vasheghani Farahani, Abbas. Ahmadi, M.H. Fazel Zarandi,, Department of Industrial Engineering and Management Systems Amirkabir University
of Technology Tehran, Iran Lung Nodule Diagnosis from CT Images bsed on ensemble learning, Research gate, August 2015, [ONLINE]. Available:
https://www.researchgate.net/publication.
[13] Sheeraz Akram, Muhammad Younus Javed, Usman Qamar, Aasia Khanum and Ali Hassan, Artificial neural network based classification of lungs nodule
using hybrid features from computerized tomographic images , Applied Mathematics & Information Sciences, 25 JUNE 2014.
[14] F.Paulin, et al; Classification of Breast cancer by comparing Back propagation training algorithms, International Journal on Computer Science and
Engineering (IJCSE).
[15] Nawaz Khana, Dhara Gaurava, Thomas Kandlb, Evaluation of Levenberg-Marquardt Technique in Error Reduction for Diabetes Condition Classification,
Elsevier, International Conference on Computational Science, ICCS 2013.
[16] Kale Vaishanw G, Morphological filtering for lung image enhancement, Proceedings of 3rd IRF International Conference, Goa, India, ISBN: 978-93-
84209-15-5., 10th May-2014.
[17] Dr .K. Vijayarekha, Feature Extraction, NPTEL Electronics & Communication Engineering Pa.ttern Recognition.
[18] Jinsa Kuruvilla, K. Gunavathi, Lung cancer classification using neural networks for CT images, Elsevier, Computer methods and programs in biomedicine,
1 1 3 (2 0 1 4 ) 202209, October 2013.
[19] Ezhil E. Nithila , S.S. Kumar, Automatic detection of solitary pulmonary nodules using swarm intelligence optimized neural networks on CT images,
Elsevier, December 2016

All rights reserved by www.ijste.org 593