2017 International Conference on Advanced Computing and Communication Systems (ICACCS -2017), Jan.

06 – 07, 2017, Coimbatore, INDIA

Neural Network Model for Automated

Detection of Tuberculosis
A.Ponmani Mrs.R.Beaulah Jeyavathana M.E, Dr.R. Balasubramanian,
Department of CSE, Assistant Professor, Professor and Head
Mepco Schlenk Engineering College, Department of CSE, Department of CSE,
Sivakasi, India. Mepco Schlenk Engineering Manonmaniam Sundaranar
[email protected] College,Sivakasi, India. University, Tirunelveli.
[email protected] [email protected]

Osman(2012) extracted geometrical features and

ABSTRACT: affine moment invariants from tissue image. To
Tuberculosis is a major health disease in many hybrid multi-layered perceptron network by these
regions of the world. In this work we suggest an
feature and to discriminate between Tuberculosis
automated approach to diagnosis the tuberculosis
bacilli and non-bacilli [4].Sputum smear
using CT scan to improve the efficiency. The
microscopy has plays an important role in TB
disease can be diagnosed easily by radiologists
detection and it has several advantages compared
with the help of automated tuberculosis detection to others [5] . But manual screening process for
system. Four stages are being used to detect screening of negative slides is more time
tuberculosis disease. They are pre-processing the consuming and labour-intensive.
image, segmentation, Extracting the feature and S. Jaeger, A. Karargyris, S. Antani, and G. Thoma
finally classification. These stages are used in introduced [6] ”Diagnosis the Tuberculosis in
medical image processing to enhance the TB radiograph using lung mask”. Combined a lung
identification. In segmentation module we use shape model, intensity model and log Gabor mask
active contour method. In feature extraction stage for segmentation. This work is not give the perfect
the Fourier-based shape descriptor technique. In accuracy.
classification stage Support Vector Neural S. Candemir, L. Folio, J. Sielgelman, [7] proposed
network is used. Automated detection of tuberculosis screening
using chest radiographs. To extract the lung region
Keywords: Tuberculosis, Active contour, Fourier using Graph cut segmentation, then compute a set
Descriptor, neural network of texture and shape features, which enable the X-
rays to be classified as normal or abnormal using a
INTRODUCTION: binary classifier.
Even though several effective methods have been
Tuberculosis is one of the infectious disease caused used to reduce the effect of TB, but it is a third high
by the bacillus Mycobacterium Tuberculosis rated disease causing death every year since just X-
(MTB). It usually affects the lungs. Tuberculosis is rays are used for detection process. X-ray is not
spreading disease through the air when people who easily predicting the early stage of tuberculosis.
have active TB in their lungs, cough, speak and Manually the detection of TB is done by just
sneeze. TB affects all age groups and all parts of looking at the X-rays by the doctors/technicians. So
the world. However, the disease mostly affects by means of looking at the images by the naked eye
young adults, and people living in developing there is more chance for wrong prediction of the
countries.Currently, there were 9.6 million cases of intensity of the tuberculosis. Hence, because of this
active Tuberculosis which resulted in 1.5 million wrong prediction of tuberculosis, we have use
deaths . automated detection of tuberculosis.
Knechel (2009)provided first lab test to detect the To overcome the problems in existing methods ,
active tuberculosis in a patient and also to find we have used CT lung image for detection of
abnormal in chest radiograph [1].Patil et al. (2009) tuberculosis automatically. A CT scan produces
using gray level co-occurrence technique for detailed images of organs, bones, and blood
estimated the texture feature from the active shape vessels. It can be used to diagnose the cancer, heart
model based on segmented lung image [2] and also disease, Tuberculosis, appendicitis, trauma and
finding the abnormal on a chest radiograph may infectious diseases. A CT scan analyse the large
suggest TB. This is poor sensitive method for the amount of data. Computed Tomography (CT) is
detection of Tuberculosis. efficient than X-ray.

ICACCS -2017, Jan. 06 – 07, 2017 IEEE ISBN No. 978-1-5090-4558-7 2164
 2017 International Conference on Advanced Computing and Communication Systems (ICACCS -2017), Jan. 06 – 07, 2017, Coimbatore, INDIA

Wiener2 filter preserves the edges and fine details

of lungs. It is low pass-filter. wiener2 filter uses a
II PROPOSED WORK: pixel wise adaptive Wiener method based on
statistics estimated from a local neighborhood of
The following Fig.2.1 shows the architectural each pixel. The filter size of 5*5 is selected to
design of the proposed system. avoid over smoothing of the image.2D Wiener2
filter is used for reduction of additive gaussian
white noise in images.

PREPROCESSING
wiener2 filter

SEGMENTATION
initial contour-Mean shift
active contour-GVF

Figure 2.2 Sample CT lung image

2.3 Segmentation:
Segmentation is carried out by two
methods, one is initial contour and the another one
FEATURE EXTRACTION
is active contour. In the initial contour method, We
Fourier Descriptor
used the technique known as mean shift
segmentation model . Mean shift segmentation is
defined as density estimation function. In the
active contour method the technique gradient
vector flow is used. The GVF forces are used to
derive the snake, which is modeled as a physical
CLASSIFICATION object having a resistance to both stretching and
Support Vector Neural bending, towards the boundaries of the object. We
Network have applied diffusion equations to both
components of the gradient of an image edge map.
It can be used to track the objects in temporal as
Fig 2.1 System Design well as the spatial dimensions. Using external GVF
forces, it can easily be manipulated.
2.1 Material and Methods
In the study, dataset containing lung CT 2.3.1 Mean Shift Segmentation Algorithm: This
images comprising abnormal lung and normal lung algorithm detects modes in a probability density
are taken from several patients was utilized. The function based on the Parzen Density Estimate
lung diseases are categorized by the radiologist
from the CT Image. Images are collected from f ks (a) 1 N a ai
d
ks ( )
male and female patients whose ages are ranging
from 15 to 78 years. All CT images are in size of No i 1 o
512 x 512 pixels and stored as DICOM (Digital Where is the number of -dimensional vectors
Imaging and Communications in Medicine) format a1 ,..., aN and the parameter „o’ is the window
files. radius of the kernel . Each feature vector
contains spatial information of each pixel and the
2.2 Pre-Processing corresponding colour/intensity information in the
Preprocessing is done to remove unwanted noise range domain of dimension one or more during
and it gives quality to the images at this stage image segmentation.
where filtering is done to remove noise. In our The multivariate mean shift vector in any point
proposed system we have used wiener2 filter to can be calculated as
remove noise.

ICACCS -2017, Jan. 06 – 07, 2017 IEEE ISBN No. 978-1-5090-4558-7 2165
 2017 International Conference on Advanced Computing and Communication Systems (ICACCS -2017), Jan. 06 – 07, 2017, Coimbatore, INDIA

N and it is also assumed that it has mean value as zero

hi k((h hi ) / l) and variance E
as follows
i 1
m(h) N
h g( p) 1 2
(G I )( p) / 2 2
E
k((h hi ) / l) e
i 1
2 E

f (x, y) 1 g( p)
Input: Pre-processed Image
The GVF segmentation algorithm :
Output: Suspected area in the image
Step 1. Set a seed point Input: Output of Initial Contour method
Step 2. Initialize the threshold value
Output: To Deformed shape (circle)
Step 3. To calculate the distance from the mean to
each pixel in the image Step1: To deform the shape. using a linear equation
as follows:
Step 4. Square the distance value and then check ax=b
the square distance value is lesser than the where a is representing the matrix.
b is vector field.
threshold;
then Step2: Interpolate the deformed shape using GVF
force
add those pixels with the cluster. If shape compressed too much
Step 5: To calculate a new mean in the cluster. Remove a neighbourhood pixel
Else
Step 6: If both the clusters are lesser than the Add some pixels in between randomly
threshold/2
Step3: Repeat the process from step 2 for more
Group them as one iteration and stop if there is no change in
Else shape

Redo the process from step 3 based on the Step4: Give some colour to differentiate the
new mean boundary (shape) from the Image.

Step 7: Get all the connected components (patches) 2.4 FEATURE EXTRACTION
in each cluster
In feature extraction We have used
Step 8: Compare each patch in the cluster with Boundary-based Fourier descriptor features that
neighbourhood clusters which adds votes describe the shape of the segmented CT image. The
shape features are extracted from the Fourier
for each pixel. descriptor (FD) based on centroid distance function
Step9: Finally, group the pixels which are all (CDF).

having more votes for a cluster. 2.4.1CENTROID DISTANCE FUNCTION

2.3.2 GVF
The Gradient Vector Flow (GVF) It is a one-dimensional (1-D) shape
framework is a well-designed, external force which signature calculated from shape boundary
deals with concave regions thereby detecting the coordinates. The r(t) function expresses the
object boundaries from both of the directions. distances of the boundary points from the centroid
Though it is based on distance transform, a binary
(y x , y y ) of a shape.
edge map is required; which estimates GVF
directly from the continuous gradient space. The r(t) [(a(x) y x )2 (b(x) y y )2 ]
diffusion process provides the GVF with a
measurement that is defined to be contextual and Shape signature is only dependent on the location
unequal to the distance from the closest point. This of the centroid and the points on the boundary, it is
is mainly due to the fact that there is a presence of invariant to the translation of the shape and also the
more than one boundary pixel taking part in the rotation of the shape if the starting point of the
estimation of the GVF. A continuous edge based original shape can be identified on the rotated
data space is detected by a Gaussian edge detector shape.

ICACCS -2017, Jan. 06 – 07, 2017 IEEE ISBN No. 978-1-5090-4558-7 2166
 2017 International Conference on Advanced Computing and Communication Systems (ICACCS -2017), Jan. 06 – 07, 2017, Coimbatore, INDIA

The x and y coordinates of the centroid, denoted by Output Layer - Arriving at a neuron in the output
gx and g y respectively, are given by layer, the value from each hidden layer neuron is
N 1 multiplied by a weight(wkj ) and the resulting
(ai ai 1 )(ai bi 1 ai 1bi ) weighted values are added with bias (vk )
yx i 0
producing a combined value. That value is fed into
6A
N 1 a transfer function, σ, which outputs a value ( yk ) .
(bi bi 1 )(ai bi 1 ai 1bi ) The y values are the outputs of the network.
i 0
yy
6A wi wi1 hj wkj vk yk
where A is area of the shape
xi
N 1
1
A (ai bi 1 ai 1 bi )
2 i 0

2.5 CLASSIFICATION

Artificial neural networks are called neural

networks or multi-layer perceptrons. Multilayer
perceptron is a feed forward neural network. It has
one input layer, one output layer and more than one
hidden layer between input and output layer.
Multilayer perceptron trained by Support Vector
Input Hidden Output
Neural Network (SVNN). Nodes in the input layer
represent the input data. Remaining nodes maps
inputs to the outputs by performing linear Figure 3.1: Architecture diagram for MLPNN
composition of the inputs with the nodes weights w
and bias b and applying activation function. RESULTS AND DISCUSSION:

A row of neurons is called a layer and one network The CT images used for testing and
can have multiple layers. The architecture of the training purpose for classification were collected
multilayer neural network from AARTHI SCANS & LABS at TIRUNELVELI.
We have several CT images, but we use 67 images
• This network has an input layer. The (on for my work out of which 23 images have
the left) layer that takes input from your tuberculosis and the remaining images do not have
dataset is called the visible layer tuberculosis. The segmentation of the image takes
place, in which initial contour is done before active
• One hidden layer (in the middle) with contour. For larger number of CTs to be examined,
three neurons and this initial contour method is impossible. In
computer aided system, the approach is to locate
• An output layer (on the right) with three the tuberculosis in the lung field and start the snake
neurons. evolution. The set of TB detected CT images and
non-TB CT images is tested to give an accurate
Input Layer -. We give the input image as CT lung result. Thus this technique deals with the accurate
image. The input layer distributes the values (xi ) detection of Tuberculosis.
to each of the neurons in the hidden layer. RESULTS

Hidden Layer - Arriving at a neuron in the hidden

layer, the input value(xi ) from each input neuron
is multiplied by a weight (wi ) , and the resulting
weighted values are added with bias value( wi1 )
and then producing a combined value .
That value is fed into a transfer function, σ, which
outputs a value (hj ) . The outputs from the hidden Figure 3.2 Image after Pre-processing of original
layer are distributed to the output layer. image (Preprocessed image)

ICACCS -2017, Jan. 06 – 07, 2017 IEEE ISBN No. 978-1-5090-4558-7 2167
 2017 International Conference on Advanced Computing and Communication Systems (ICACCS -2017), Jan. 06 – 07, 2017, Coimbatore, INDIA

Figure 3.3 After Initial contour

Figure 3.6 Gradient graph

Figure 3.4 Output of Active contour

NEURAL NETWORK CLASSIFICATION

RESULTS: Figure 3.7 Confusion Matrix for Training

Figure 3.8 Confusion Matrix for Testing

Figure 3.5 The best performance validation is 2.907 Conclusion and Future Work:
07 In this paper, an approach for automatic
e at epoch 45. TB detection from CTs image has been proposed.
So far, there are no automatic algorithms developed
that can detect TB from CTs accurately. In this
approach, a mean shift segmentation technique
integrated with adaptive thresholding to automate
the initial contour has been done. These initial
contours are further used in a mean shift algorithm
model to segment out suspected features and active

ICACCS -2017, Jan. 06 – 07, 2017 IEEE ISBN No. 978-1-5090-4558-7 2168
 2017 International Conference on Advanced Computing and Communication Systems (ICACCS -2017), Jan. 06 – 07, 2017, Coimbatore, INDIA

contour used in gradient vector flow (GVF). In the L. Shunmuganathan. “Smart Spotting of
subsequent classification process, neural network Pulmonary TB Cavities Using CT Images”
are applied. Experimental results demonstrate that
our method achieves good accuracy.

REFRENCES:

[1] Knechel NA. Tuberculosis: pathophysiology,

clinical features, and diagnosis. Crit Care Nurse
2009;29(2):34–43.
[2] Patil SA, Udupi VR, Kane CD, Wasif AI,
Desai JV, Jadhav AN. Geometrical and texture
features estimation of lung cancer and TB images
using chest X-ray database. Proc. of Int. Conf.
Biomedical and Pharmaceutical Engineering; 2009.
pp. 1–7.
[3] G. F. Costa, F. F. Costa Filho, F. Sena, J.
Salem, and O. de Lima, “Automatic
identification of mycobacterium tuberculosis
with conventional light microscopy,” in Annual
International Conference of the IEEE Engineering
in Medicine and Biology Society (EMBC ’08), pp.
382–385, 2008.
[4] Osman MK, Mashor MY, Jaafar H.
Performance comparison of extreme learning
machine algorithms for mycobacterium
tuberculosis detection in tissue sections. J Med
Imaging Health Inf 2012;2(3):307–12.
[5] Raof RAA, Mashor MY, Ahmad RB, Noor
SSM. Image segmentation of Ziehl-Neelsen sputum
slide images for tubercle bacilli detection. In: Ho
PG, editor. Image segmentation. InTech; 2011. p.
365–78 [chapter 19].
[6] S. Jaeger, A. Karargyris, S. Antani, and G.
Thoma, “Detecting tuberculosis in radiographs
using combined lung masks,” in Proc. Int. Conf.
IEEE Eng. Med. Biol. Soc., 2012, pp. 4978–4981.
[7] S. Jaeger, A. Karargyris, S. Candemir, L. Folio,
J. Sielgelman, F.Callaghan, Z. Xue, K
.Palaniappan, R. Singh, S. Antani, G. Thoma,Y.-X.
Xiang, P.-X. Lu, and C. McDonald, “Automatic
tuberculosis screening using chest radiographs,”
IEEE Trans. Medical Imaging,2013, to be
published.
[8] Lung tuberculosis identification based on
statistical feature of Thoracic X-ray
Ratnasari Nur Rohmah;Adhi Susanto; Indah
SoesantiQiR (Quality in Research), 2013
International Conference on year: 2013.
[9] ISSN: 2278 –909X International Journal of
Advanced Research in Electronics and
Communication Engineering (IJARECE) Volume4,
Issue 6, June 2015.
[10] P. C. Hopewell, M. Pai, D. Maher, M.
Uplekar, and M. C. Raviglione, “International
standards for tuberculosis care,” Lancet Infectious
Diseases, vol. 6, no. 11, pp. 710–725, 2006.
[11] V. Ezhil Swanly, L. Selvam,P. Mohan
Kumar,J. Arokia Renjith, M. Arunachalam,4 and
K.

ICACCS -2017, Jan. 06 – 07, 2017 IEEE ISBN No. 978-1-5090-4558-7 2169