Statistical Feature-based Neural Network Approach for the Detection of Lung Cancer

K.A.G. Udeshani, R.G.N. Meegama & T.G.I. Fernando
International Journal of Image Processing (IJIP), Volume (5) : Issue (4) : 2011 425
Statistical Feature-based Neural Network Approach for the
Detection of Lung Cancer in Chest X-Ray Images
K.A.G. Udeshani gayaudeshani@gmail.com
Excel Technology Lanka Ltd
Thimbirigasyaya Road
Colombo 05, Sri Lanka
R.G.N. Meegama rgn@dscs.sjp.ac.lk
Department of Statistics and Computer Science
Faculty of Applied Science
University of Sri Jayewardenepura
Gangodawila, Nugegoda, Sri Lanka
T.G.I. Fernando gishantha@dscs.sjp.ac.lk
Department of Statistics and Computer Science
Faculty of Applied Science
University of Sri Jayewardenepura
Gangodawila, Nugegoda, Sri Lanka
Abstract
Lung cancer, if detected successfully at early stages, enables many treatment options,
reduced risk of invasive surgery and increased survival rate. This paper presents a novel
approach to detect lung cancer from raw chest X-ray images. At the first stage, we use a
pipeline of image processing routines to remove noise and segment the lung from other
anatomical structures in the chest X-ray and extract regions that exhibit shape characteristics
of lung nodules. Subsequently, first and second order statistical texture features are
considered as the inputs to train a neural network to verify whether a region extracted in the
first stage is a nodule or not . The proposed approach detected nodules in the diseased area
of the lung with an accuracy of 96% using the pixel-based technique while the feature-based
technique produced an accuracy of 88%.
Keywords: Lung Nodule, Computer Assisted Diagnostic, Artificial Neural Network, Chest
Radiography, Medical Imaging
1. INTRODUCTION
Lung cancer is the growth of a tumor, referred to as a nodule, that arise from cells lining the
airways of the respiratory system. These cells are often in bright contrast in chest X-rays and
take the shape of a round object. However, these nodules that can be seen in a chest X-ray
may not necessarily be a lung cancer; it can be due to some other disease such as
pneumonia, tuberculosis or calcified granuloma. As such, the detection of lung cancer has
been a tedious task in medical image analysis over the past few decades. If lung nodules can
be identified accurately at an early stage, the survival rate of the patients can be increased by
a significant percentage. In the health industry, chest X-rays are considered to be the most
widely used technique for the detection of lung cancer. However, because it is difficult to
identify lung nodules using raw chest X-ray images, analysis of such medical images has
become a tedious and complicated task. This paper presents a novel technique that can be
used to detect lung cancer in early stages.
With the evolution of various technologies, several projects have been carried out to identify
lung cancer at early stages. In these projects, digital photographs of chest X-rays and CT
scans have been used to identify lung cancer. In recent research literature, it is observed that
principles of neural networks have been widely used for the detection of lung cancer in
medical images with simulated lung nodules [1, 5], massive training artificial neural networks
[11], two level neural classifiers [23], hybrid lung nodule detection [24], and ladder structured
decision trees [6, 25]. A co-occurrence matrix and texture measures have been used in [3] to

detect lung nodules. A rectangular area of a chest X-ray was considered in [4] as the input to
a neural network.
Apart from using chest X-rays, researchers have used an ensemble of artificial neural
networks to detect cancer cells in specimen images of needle biopsies [2]. A template
matching algorithm [7] and genetic algorithms [8] have been used to detect lung nodules.
These are quite fundamental approaches to detect nodules where the similarity between an
unknown object (observed image) and a template (reference image) are compared.
One of the most common problems encountered during identification of lung nodules is
overlapping of the ribs and clavicles with nodules. Such overlapping of images results in the
detection of suspicious areas difficult. Several rib suppression techniques have been
proposed in recent research literature such as blind source separation [9], texture analysis
[12], watershed segmentation [10], Gaussian filters [14], active shape modeling [15] and
quasi-gabo filters [12].
In [13], a set of features that contain translation invariant wavelet based features and co-
occurrence features of mammograms were used to classify images. Features extracted from
a multi scale Gaussian filter bank and some specific features that are readily calculated from
the blob detector scheme have been used to detect nodules [15]. The local curvature of the
image data were considered when viewed as a relief map [22].
An image subtraction approach was proposed in [16] where a test image and a comparison
image were used. After applying several widely used pre-processing techniques, the two
images were subtracted to obtain a resultant image which can be used to identify cancers
with candidate nodules. This research, however, did not verify the utility of the study using
many chest X-ray films.
A heuristic approach using artificial intelligence tools have been used in [26] to detect cancer
nodules.
2. METHODOLOGY
2.1 Overview
The proposed algorithm consists of two stages as depicted in Fig. 1. The first stage involves a
pipeline of image processing routines to separate the lung from other structures of original
chest X-rays and identifying the region that is suspected of being a nodule. At this stage, the
system extracts 65x65 square areas considering the suspicious point as the center. Because
it involves a pixel-based technique, all the pixels in the square region are considered as the
inputs to the system. The intensity values of the pixels that fall within this region are extracted
and are stored in a database which is used to train the system at the next stage. The
database is then divided into several sub categories, and the data available in these sub
categories will be used for the training as well as for testing the results.
In the second stage, a neural network is trained based on two types of inputs: pixel-based
inputs in which we consider the intensity levels of pixels within the suspected region and
statistical feature-based inputs where we take into account first and second order statistical
features.

FIGURE. 1: Stages of the proposed algorithm.
(a) (b) (c)
FIGURE. 2. Pre-processing of chest X-rays. (a) original image, (b) median filtering and (c) histogram
equalization
2.2 Pre-processing
As shown in Fig. 2, Median filtering is required to remove the effect of poor contrast due to
glare, noise and effects caused by poor lighting conditions during image capture. A low-
frequency image was generated by replacing the pixel value with a median pixel value
computed over a square area of 5x5 pixels centered at the pixel location. Sharpening and
histogram equalization methods were used to enhance the contrast of the images.

(a) (b)
FIGURE. 3. Lung separation. (a) binary thresholded image and (b) suspicious region.
The image was then converted to binary image by using a thresholding technique. It
computes a global threshold that can be used to convert an intensity image to a binary image.
It utilizes Otsu's method, which chooses the threshold to minimize the intra-class variance of
black and white pixels [17].
2.3 Lung Region Segmentation
Lung masks were prepared using active shape models that are available with the JSRT
database [27]. These images can be used to segment the lung region while the user can
identify the scope when selecting suspicious points [15]. These images consist of 154 lung
nodules (100 malignant cases, 54 benign cases), and 93 non-nodules. The chest radiographs
were digitized with a 0.175mm pixel size, a matrix size of 2048 x 2048 and a 12-bit grayscale
level. The criteria for inclusion of radiographs in the database were: (1) absence of nodules
larger than 35 mm, (2) absence of suspicious nodules that were not confirmed by CT
examination, (3) no more than one nodule per patient, and (4) absence of nodules with
margins that could not be confirmed by radiologists. The subtlety of the images contained in
this database are grouped into five categories, namely, obvious, relatively obvious, subtle,
very subtle and extremely subtle. These categories have been defined by expert radiologists
which takes into account the size, contrast, and anatomical position of the lesions.
We used connected components labeling to scan the binary thresholded image and group its
pixels into components based on pixel connectivity, i.e.. all the pixels in a connected
component share similar pixel intensity values and are in some way connected with each
other as in Fig. 3 [17].
We define the roundness index of each connected component Ci as
2
4
i
i
i
R
A
C
π
=
where Ai is the area, Ri is the perimeter of the i
th
component. This index, that gives a measure
of irregularity of a connected component, is equal to 1 only for circles and less than 1 for any
other shape. As such, if a connected component exhibits a roundness index closer to 1, the
probability of it being a nodule can be considered to be high.
After identifying a region that shows a high probability of being a nodule, we proceed to the
second stage of our algorithm to train the neural network.

FIGURE. 4: Gray level co-occurrence matrix with different offsets. Shaded region at the middle indicates
the pixel of interest.
Six features based on first order statistical textures measured with the gray level count of the
image and four second order statistical texture features based on how often pairs of pixel with
specific values and in a specified spatial relationship occur in an image were considered as
inputs to a neural network [18, 19]. Average gray level, standard deviation, smoothness of
intensity in a region, skewness (third moment) of the histogram, uniformity and entropy were
the first order statistical texture features that we considered. First order statistical features
consider statistical moments of the region of interest where the n
th
statistical moment is given
by
∑
−
=
−=
1
0
)()(
L
i
i
n
in zpmzµ
where L is the possible number of intensity levels, m is the mean of intensity values, zi is the
intensity value and p(zi) is the percentage of pixels with intensity value zi. Thus, skewness is
considered to be the 3
rd
statistical moment. The entropy e, which provides a measure of
randomness in the pixels, is given as
)(log)( 2
1
0
i
L
i
i zpzpe ∑
−
=
−=
The second order statistical texture features measure: (a) the local variations in the gray level
co-occurrence matrix (GLCM), (b) correlation that measures the joint probability occurrence of
the specified pixel pairs, (c) energy that provides the sum of squared elements in the GLCM
and (d) homogeneity that measures the closeness of the distribution of elements in the GLCM
to the GLCM diagonal.
The GLCM, also known as the gray-level spatial dependence matrix, is a statistical method of
examining texture and considers the spatial relationship between pixels. First, we create a
GLCM with different offsets, expressed as an angle, as seen in see Fig. 4. To create this
GLCM, we consider the above directions which specify the offset values with the common
angles and the pixel distance. A 1-by-2 array of integers specifies the distance between the
pixel of interest and its neighbor. Subsequently, we extract the statistical measures from
these matrices as shown in Fig. 5 to obtain an average.
A neural network with one hidden layer of 1000 neurons and an input layer of 10 neurons to
accommodate the ten first and second order textures was used at the training stage [19, 21].

FIGURE. 5: Creating the GLCM matrix.
Suspicious region
Recognition rate (%)
Pixel-based Feature-based
Nodule 96 88
Non-nodule 88 7
TABLE 1: Recognition rates for the two types of inputs to the neural network.
3. RESULTS AND DISCUSSION
A 49x49 mask was selected within a suspicious region via an interactive user interface that
we have developed for this purpose. After selecting the suspicious region during the first
stage of our algorithm, we train the neural network until convergence is achieved. Two types
of inputs were considered as inputs to the neural network. One type of inputs consists of
intensity levels of pixels within the 49x49 mask. The other contains the ten statistical texture
features within this mask as described in Section 2.3.
This system uses a neural network with one hidden layer containing 1000 neurons, and an
output layer with 1 neuron as in Fig. 7. For the pixel-based intensity input vectors, we used
purelin and tansig transfer functions whereas two tansig transfer functions were used for the
statistical feature-based inputs vectors. Backpropagation was created by generalizing the
Widrow-Hoff learning rule to multiple-layer networks and non-linear differentiable transfer
functions. Fig. 8 demonstrates the performance of the neural network for these two types of
inputs considering the training approaches and the mean squared error (MSE). It was
observed that during pixel-based technique, the R value was 1 and the MSE was 1.2682e-
009. The feature-based training resulted in an R value of 0.737 and an MSE of 0.2684.
As seen in Table 1, we managed to achieve a high recognition rate for a nodule when the
neural network was trained using pixel-based intensity values. Recognizing a non-nodule was
16% lower with statistical feature-based training of the neural network. We have used a
pipeline of image processing routines at the first stage to identify regions suspected to be of
lung nodules from the visible regions of chest X-rays. The second stage involves a neural
network trained using backpropagation to verify whether the suspected region is a nodule or
not. The reason for applying a neural network for the verification is due to detection of false-
positives during the image processing stage.

FIGURE. 6. Selecting a suspicious region on the image to determine statistical features or consider pixel
intensity values.
FIGURE 7: Architecture of the neural network.
The selection of suspicious region involves a 49x49 square mask which include nodules with
13 mm diameter with a resolution of 96 pixels per inch. In the JSRT database, the size of a
nodule varies between 8.9 mm and 29.1 mm with an average size of 17.4 mm. Because in
medical imaging literature, the size of a lung nodule if detected at an initial stage, is
considered to be between 5mm and 20 mm.
We have used the masking method for lung region segmentation as the JSRT database
provides all the required masks for the chest X-rays. As the final results depends on the
subtlety of the images, only images with a low subtlety have been used.
In contrast to other computer-aided lung cancer detection methodologies identified in
research literature that are based on artificial neural networks, the proposed network uses 10
features as inputs to the system.

(a) (b)
FIGURE 8: Training graph of the neural network: (a) pixel-based and (b) statistical feature-based inputs.
4. CONCLUSION
In this research, we have successfully developed a solution for the detection of lung cancer
nodules using image processing algorithms and neural networks. The system integrates a
graphical user interface where the user is able to select a digital chest X-ray image as the
input and the system will show suspected regions of the image that contain lung nodules.
The detection of nodules is confounded by overlapping shadows of blood vessels and ribs.
Although researchers have proposed several techniques to reduce the effect of ribs in chest
X-rays to eliminate false-positives, it is prudent to consider hidden regions too, such as
behind the heart, spine and the diaphragm, and apply relevant filtering techniques to
suppress these anatomical structures.
5. REFERENCES
[1] P.R. Snoeren, G.J.S. Litjens, B.V. Ginneken and N. Karssemeijer, Training a computer
aided detection system with simulated lung nodules in chest radiographs, Proc. 3rd
International Workshop on Pulmonary Image Analysis, Beijing, 2010.
[2] Z.H. Zhou, Y. Jiang, Y.B. Yang and S.F. Chen, Lung cancer cell identification based on
artificial neural network ensembles, Artificial Intelligence in Medicine, vol. 24, N 1, pp. 25-
36, 2002.
[3] G.S. Cox, F.J. Hoare and G. de Jager, Experiments in lung cancer nodule detection using
texture analysis and neural network classifiers, Third South African Workshop on Pattern
Recognition, 1992.
[4] J.W. Lee, H.W. Lee, J.H. Lee, I.T. Kang and G.K. Lee, A study on lung nodule detection
using neural networks, Proc. IEEE Region 10 Conference, pp. 1150 - 1153, 1999.
[5] G. Coppini, S. Diciotti, M. Falchini, N. Villari and G. Valli, Neural networks for computer
aided diagnosis: detection of lung nodules in chest radiograms, IEEE Trans. on
Information Technology in Biomedicine, vol. 4, pp. 344-357, 2003.
[6] M.G. Penedo, M.J. Carreira, A. Mosquera and D. Cabello, Computer aided diagnosis: A
neural network based approach to lung nodule detection, IEEE Trans. on Medical
Imaging, vol. 17, N 6. pp. 872-880, 1998.
[7] M.N. Gurcan, B. Sahiner, N. Petrick, H.P. Chan, E.A. Kazerooni, P.N. Cascade and L.
Hadjiiski, Lung nodule detection on thoracic computed tomography images: Preliminary
evaluation of a computer-aided diagnosis system, Medical Physics, vol. 29, N 11, pp.
2552, 2002.

[8] Y. Lee, T. Hara, H. Fujita, S. Itoh and T. Ishigaki, Nodule detection on chest helical CT
scans by using a genetic algorithm, Proc. Intelligent Information Systems, pp. 67 - 70,
1997.
[9] T. Rasheed, B. Ahmad, M. A. U. Khan, M. Bettayeb, S. Lee and T. S. Kim, Rib
suppression in frontal chest radiographs: A blind source separation approach, Proc.
Signal Processing and its Applications, pp. 1 - 4, 2007.
[10] K. Le, Automated detection of early lung cancer and tuberculosis based on X-ray image
analysis, Proc. WSEAS International Conference on Signal, Speech and Image
Processing, pp. 1-6, 2006.
[11] K. Suzuki, J. Shiraishi, H. Abe, H. MacMahon and K. Doi, False-positive reduction in
computer-aided diagnostic scheme for detecting nodules in chest radiographs by means
of massive training artificial neural network, Academic Radiology, vol. 12, N 2, pp. 191-
201, 2003.
[12] M. Park, J. S. Jin and L. S. Wilson, Detection of abnormal texture in chest X-rays with
reduction of ribs, Proc. Pan-Sydney area workshop on visual information processing,
2004.
[13] L. Zhang and X. Gao, Research on translation invariant wavelet transform for
classification in mammograms, Proc. 3rd international conference on natural
computation, 2007.
[14] P. Campadelli, E. Casiraghi and D. Artioli, A fully automated method for lung nodule
detection from postero-anterior chest radiographs, IEEE Trans. on Medical Imaging, vol.
25, N 12, pp. 1588-1602, 2006.
[15] A.M.R. Schilham, B.V. Ginneken and M. Loog, A computer-aided diagnosis system for
detection of lung nodules in chest radiographs with an evaluation on a public database,
Medical Image Analysis, vol. 10, N 2, pp. 247-258, 2006.
[16] R. Hayashibe, N. Asano, H. Hirohata, K. Okumura, S. Kondo, S. Handa, M. Takizawa, S.
Sone and S. Oshita, An automatic lung cancer detection from X-ray images obtained
through yearly serial mass survey, Proc. International conference on image processing,
pp. 343 - 346, 1996.
[17] R. C. Gonzalez , R.E. Woods and S. L. Eddins. Digital Image Processing, Addison-
Wesley, 2002.
[18] S. A. Patil and V. R. Udupi, Chest X-rays features extraction for lung cancer
classification, Journal of Scientific and Industrial Research, vol. 69, pp. 271-277, 2010.
[19] J. Wei, Y. Hagihara, A. Shimizu and H. Kobatake, Optimal image feature set for detecting
lung nodules on chest X-ray images, Proc. Int. Workshop on Computer-Aided
Diagnosis, 2002.
[20] O. Ludwig and U. Nunes, Novel maximum- margin training algorithms for supervised
neural networks, IEEE Trans. on Neural Networks, vol. 21, N 6, pp. 972-983, 2010.
[21] R. N. Strickland, Tumor detection in non-stationary backgrounds, IEEE Trans. on Medical
Imaging, vol. 13, pp. 491-499, 1994.
[22] S.B. Lo, S.L. Lou, J.S. Lin, M.T. Freedman and S.K. Mun, Artificial convolution neural
network techniques and applications for lung nodule detection, IEEE Trans. on Medical
Imaging, vol. 14, pp. 711-718, 1995.

[23] J.S. Lin, S.B. Lo, A. Hasegawa, M.T. Freedman and S.K. Mun, Reduction of false
positives in lung nodule detection using a two-level neural classification, IEEE Trans. on
Medical Imaging, vol. 15, pp. 206-216, 1996.
[24] Y.S.P. Chiou, Y.M.F. Lure and P.A. Ligomenides, Neural networks image analysis and
classification in hybrid lung nodule detection (HLND) system, IEEE Workshop on
Neural Networks for Signal Processing, pp. 517-526, 1993.
[25] D.H. Ballard and J. Sklansky, A ladder-structured decision tree for recognizing tumors in
chest radiographs, IEEE Trans. on Computers, vol. C-25, pp. 503-513, 1976.
[26] B.V. Ginneken, B.M.T.H. Romeny and M.A. Viergever, Computer Aided Diagnosis in
chest radiography: A survey, IEEE Trans. on Medical Imaging, vol. 20, pp. 1228-1241,
2001.
[27] Japanese Society of Radiological Technology, http://www.jsrt.or.jp

Statistical Feature-based Neural Network Approach for the Detection of Lung Cancer

More Related Content

What's hot (20)

Similar to Statistical Feature-based Neural Network Approach for the Detection of Lung Cancer (20)

Recently uploaded (20)

Statistical Feature-based Neural Network Approach for the Detection of Lung Cancer