Early studies of lung tumor detection and

diagnosis using neural network

backpropagation and fuzzy
logic systems

Tumpal Pandiangan1, Ika Bali2, and A.R.J. Silalahi1

1 Medical Physics Study Programme, Faculty of Science, Technology, Engineering and
Mathematics, Matana University, jl. CBD Barat Kav.1, Tangerang 15810, Indonesia
2 Institute of Research, Science Development and Technology Study, Matana University,

Tangerang 15810, Indonesia

E-mail: [email protected]
ROAD MAP
2020/2021
Application of machine
2019/2020 learning
Development of
enhancement and
segmentation using
2018/19 machine learning for
Development of enhancement identification lung
and segmentation techniques nodule of Indonesian
2017/18 using machine learning for database.
Development of object identification lung
enhancement and nodule base on Japan
segmentation techniques database.
for lung nodule
identification base on
Japan database.
Background
 Most of those who are positively diagnosed are patients
who are 65 years or older, with only less than 2% are
younger than 45. In Indonesia, the figures show similar
trends with thousands of newly diagnosed every year.

3
 US Cancer Cases and Deaths-
2017 Estimates
Lung & bronchus, estmated Prostate , estmated
deaths 7 from 10 male deaths 2 from 10 male
Background
• Normally the lung cancer detection is done by using chest x-
ray to look for any abnormal areas in the lungs.
• If this plain x-ray detection reveals something suspicious, the
doctor will order more tests such as Computed Tomography
(CT) scan to reveal the 3 dimensional structure of the
suspected object.
• The normal x-ray procedures come with limitations because
images collected are mixed with noises, thus make it
challenging to correctly interpret the result.
Objective
To obtain a result of identification of
early lung tumors using backpropagation
neural networks and fuzzy logic systems
to assist radiologists in diagnosing
patient illnesses.
Materials and methods

Samples : Lung x-ray image

Standard chest x-ray image with image parameters are:
matrix size 2048 x 2048, pixel size 0.175 mm, number of
gray levels 4096 (12bit) and has been known severity as
maglinant at upper right of lung,
This image is taken From a standard digital image
database, the community radiology technology of
Japanese society

8
Methods :
Original
image
Gray scale
image
Enhancement :
Filtering in the spatial domain and frequency
domain

Segmentation process :
Threshold
Morphology
Edge detection

Object separation , characterization and optimization of

physically characteristic of object
Neural Network
Fuzzy Logic

9
Methods :
Digital image processing :
1. RGB conversion to gray level
2. Enhancement through spatial filtration and frequency;
3. Segmentation through thresholding
4. Morphological process
5. Identification and diagnosis using backpropagation
neural networks and fuzzy logic systems.
Figure 1. Original image convert to grayscale Image
11
Figure 2. Grayscale Image convert to the product of filtering
in frequency and spatial domain
12
Figure 3. Product of segmentation process on image of
product filtering
13
Figure 4. Image product of (a) erosion processed and (b) dilation
processed
14
Figure 5. Object location on original image
Figure 6. Image product of edge detection after dilation process
Figure 7. Selected image after dilation process
17
Image processing to get relevant and important
variables:
Perimeter
Area
Shape ratio/irregularity
 𝒑 𝟏
𝒖𝒌 = 𝒋=𝟏 𝑾𝒌𝒋.𝑿𝒋 𝒀𝒌 = 𝟏+𝐞𝐱𝐩(−𝒖𝒌)

Rules of Delta Wnew=Wold +βEX/IXI

Figure 1. Learning rules of neural networks backpropagation14
Figure 8. The process of identification and classification by artificial neural networks
Results and discussion

Results of images processing with enhancement,

segmentation and feature separation of each of
45 standard x-ray lung samples, 25 samples are
standard lung nodule (LN) and the remaining 20
samples are standard non-lung nodule (NLN).
 LN1 LN2 LN3 LN4

LN5 LN6 LN7 LN8

Figure 2. Example of lung nodule image (LN1 to LN8) after enhancement process, segmentation and separation feature14
22
Table 1. Data of area, perimeter and shape for lung nodule (LN)
Perimeter Diameter
Code Area/pixel Shape Stage
/pixel [cm]
LN1 1100 130.089 0.816396575 0.655087878 First Stage
LN2 7318 99.515 0.927107131 1.689660389 First Stage
LN3 467 79.964 0.917312896 0.426836609 First Stage
LN4 920 114.831 0.816396575 0.599096985 First Stage
LN5 608 88.755 0.969410149 0.487029211 First Stage
LN6 476 79.212 0.952828271 0.430929968 First Stage
LN7 1100 130.089 0.816396575 0.655087878 First Stage
LN8 733 98.863 0.941946082 0.534755435 First Stage
LN9 470 79.311 0.938470544 0.428205409 First Stage
LN10 622 89.144 0.983095683 0.492604543 First Stage
LN11 469 79.311 0.936473798 0.427749629 First Stage
LN12 731 99.515 0.927107131 0.534025394 First Stage
LN13 467 79.464 0.928892968 0.426836609 First Stage
LN14 1128 133.273 0.797653713 0.663372968 First Stage
LN15 920 114.831 0.876312836 0.599096985 First Stage
LN16 608 88.755 0.969410149 0.487029211 First Stage
LN17 851 105.361 0.962851767 0.57619303 First Stage
LN18 1154 123.241 0.954300214 0.670974669 First Stage
LN19 1193 123.308 0.985479457 0.682218421 First Stage
LN20 1154 123.241 0.954300214 0.670974669 First Stage
Note: First stage = (1-5) cm, Second stage = (5.1-7.0) cm, Third stage = (> 7.1) cm.
Table 2. Data of area, perimeter and shape for test sample lung nodule
Code Area Perimeter Shape
LNsu1 95 35.526 0.945410936
LNsu2 412 73.016 0.970623452
LNsu3 89 34.418 0.94364444
LNsu4 92 35.972 0.892993499
LNsu5 133 40.894 0.998900188
Figure 10. The training results for the process output approach the target
value14
Table 3. Results of ANN test on test samples (LNsu)

Results LNsu 1 LNsu 2 LNsu 3 LNsu 4 LNsu 5

Test result 0.99999 0.99998 0.99999 0.96432 0.9993

Targeted training 1 1 1 1 1
The calculation 0.945410936 0.970623452 0.94364444 0.892993499 0.99890018
results
Table 4. Data of area, perimeter and shape for test sample non lung nodule (LNsu)
Code Area Perimeter Shape
NLNsu1 436 91.03 0.660856313
NLNsu2 297 94.049 0.421733135
NLNsu3 389 85.778 0.664029554
NLNsu4 259 64.83 0.773993583
NLNsu5 415 79.014 0.521052288
Table 5. The fuzzy set of input and output variables
Membership Range
Functions Variable Domain
Functions
Input1 Shape Low 0.00-0.33
Medium 0.00-1.00 0.34-0.66
High 0.67-1.00
Input2 Object Low 0.00-5.10
Diameter Medium 0.00-10.00 5.20-7.10
High 7.20.10.00
Output Results of Low 0.00-0.33
diagnosis Medium 0.00-1.00 0.34-0.66
High 0.67-1.00
Figure 11. Results of the program mapping the relation of two inputs to one output

30
Figure 12. Results of the two-input mapping program with one output in
three-dimensional space
Table 6. Results of fuzzy logic application program for lung nodule diagnosis
Object Object Results of
Shape Diameter Diagnosis
0.82 0.70 0.35
0.93 1,7 0.50
0.92 0.43 0.17
0.82 0.60 0.22
0.97 0.44 0.17

32
Conclusion
This study have obtained a result of identification of early lung tumors using
backpropagation neural networks and fuzzy logic systems of 5 identified
samples as malignant type with condition in early stages.
The application program can generate a method of identifying lung nodule
and diagnostic results, after x-ray lung nodule image undergoes digital image
processing, neural network process and fuzzy logic inference system.
This process can produce data to assist radiologists in the treatment or
treatment of patients with early lung cancer.
ACKNOWLEDGEMENTS

The authors express their appreciation to

Matana university for the support of this
research and the research publication.

34
References
1. E.E. Coche, B. Ghaye, J. de Mey, and P. Duyck (Eds.), Comparative interpretation of CT and standard radiography of the chest, Springer-Verlag Berlin
Heidelberg, 27- 49 (2011).

2. T. Pandiangan and I. Bali, “Development of enhancement and segmentation techniques to detect early tumors in the lungs,” Journal of Medical Physics
and Biophysics, (accepted for publication) (2017) (in Bahasa).

3. P. Sprawls, “Optimizing medical image contrast, detail and noise in the digital era,” Medical Physics International Journal, 2(1), 41-48 (2014).

4. N. Ganesan, K. Venkatesh , M.A. Rama and A. Malani, “Application of neural network in diagnosing cancer disease using demographic data,”
International Journal of Computer Application, 1(26), 76-85 (2010).

5. D.S. Rao, “Comparison of fuzzy and neuro fuzzy image fusion techniques and its applications,” International Journal of Computer Applications, 43(20),
3-37( 2012).

6. N.Y. Suple and S.M. Kharad, “Basic approach to image contrast enhancement with fuzzy inference system,” International Journal of Scientific and
Research Publications, 3(6), 1-4 (2013).

7. A.M. Hamad, “Lung cancer diagnosis by using fuzzy logic,” IJCSMC, 5(3), 32- 41 (2016).

8. A. Dhary, ”Lung cancer detection using digital image processing” IJREAS 2(2), 1351-1359 (2012).

9. M.S. Al-Tarawneh, “Lung cancer detection using image processing techniques,” Leonardo Electronic Journal of Practices and Technologies, 20, 147-
158 (2012).

10. L. Tseng and L. Huang, ”An adaptive thresholding method for automatic lung segmentation in CT images,” IEEE AFRICON, 1-5, (2009).

11. F. Taher and R. Sammouda, “Lung cancer detection by using artificial neural network and fuzzy clustering methods,” IEEE GCC Conference and
exhibition, 295-298 (2011).

12. J. Tong, D.Z. Zhao, W. Ying, X.H. Zhu, and W. Xu, “Computer-aided lung nodule detection based on CT images,” IEEE-International Conference on
Complex Medical Engineering, 816-819 (2007).

13. N.S. Lingayat and M.R. Tarambale, “A computer based feature extraction of lung nodule in chest x-ray image,” International Journal of Biochemestry
and Bioinformatics, 3(6), 624-629 (2013).

14. T. Pandiangan, I. Bali and ARJ. Silalahi, “Lung cancer detection using artificial neural network,” Journal of Atom Indonesia, (Submission process)
(2018).

35