Using Distance Measure based Classification in Automatic Extraction of Lungs Cancer Nodules for Computer Aided Diagnosis

Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.3, June 2021
DOI: 10.5121/sipij.2021.12303 25
USING DISTANCE MEASURE BASED CLASSIFICATION
IN AUTOMATIC EXTRACTION OF LUNGS CANCER
NODULES FOR COMPUTER AIDED DIAGNOSIS
Maan Ammar1
, Muhammad Shamdeen 2
, MazenKasedeh 2
, Kinan Mansour3
, and
Waad Ammar3
1
AL Andalus University for Medical Sciences, Faculty of Biomed. Eng.,
Al Kadmous, Syria
2
Damascus University, Department of Biomedical Engineering, Syria
3
Al Andalus University Hospital, Al Kadmous, Syria
ABSTRACT
We introduce in this paper a reliable method for automatic extraction of lungs nodules from CT chest
images and shed the light on the details of using the Weighted Euclidean Distance (WED) for classifying
lungs connected components into nodule and not-nodule. We explain also using Connected Component
Labeling (CCL) in an effective and flexible method for extraction of lungs area from chest CT images with
a wide variety of shapes and sizes. This lungs extraction method makes use of, as well as CCL, some
morphological operations. Our tests have shown that the performance of the introduce method is high.
Finally, in order to check whether the method works correctly or not for healthy and patient CT images, we
tested the method by some images of healthy persons and demonstrated that the overall performance of the
method is satisfactory.
KEYWORDS
Nodules classification, lungs cancer, morphological operators, weighted Euclidean distance, nodules
extraction
1. INTRODUCTION
According to World Health Organization (WHO) statistics, lung cancer cases reached in 2012
one million and eight hundred thousand (13%) of all cancer cases appeared in the world,
surpassing breast cancer which came second with 11.9%[1]. Moreover, according to WHO
website, there were 1.37 million death cases due to lung cancer in 2008, which means 18% of all
death cases due to all cancer types. These facts made lung cancer a major concern for both
specialists and scientists. It is well known that early diagnosis can improve the effectiveness of
treatment and increase the patient's chance of survival[2]. The previous facts motivated
researchers to pay a great attention to researches that work on automated diagnosis of lung cancer
in a wide field known as Computer Aided Diagnosis Systems for Lung Cancer. A reliable
computer diagnosis of the disease will help screening a large number of images created every day
enabling specialized doctors to work with only little amount of candidate images and raising their
efficiency[3].
The great attention paid to research on this subject led to more than 300 published scientific
papers during the past 3 decades[2]. It motivated also several groups of researchers to write
review papers that evaluate the overall situation of research on this subject, identify the
challenges, and propose what is needed to improve the performance of the CAD approaches for

26
lung cancer detection and diagnosis. Heang-Ping Chan et al[3], wrote in 2008 a review paper on
Computer Aided diagnosis of Lung Cancer using 158 references spanning 3 decades and finished
to the conclusion that the developments in this area are still in early stage. Firmino et al.[4],
published in 2014 another review paper on " Computer-aided detection system for lung cancer in
computed tomography scans" using 77 references selected from 420 papers found in the famous
related databases, and concluded that many, if not all systems, described in their survey have the
potential to be important in clinical practice, but further research is needed to improve existing
systems and propose new solutions.Bhavanishankar K. et al. [5], wrote a survey paper in 2015
using 68 references spanning almost the previous 2 decades, on "techniques for detection of
solitary pulmonary nodules in human lung and their classifications" in an attempt to summarize
various methods that have been proposed by several authors over the years of their research on
detection, classification and diagnosis of lung nodules.In 2010, M.V. Sprindzuk, et al.[6],
published a survey on Lung cancer differential diagnosis of pulmonary nodules detection using
101 references published in the period 2004-2009 and discussed several issues including early
diagnosis, and reached an optimistic view of the ability of using the developed systems for
different modalities of images, and called for improving the techniques of the image analysis to
increase the sensitivity of diagnostic strategies. Finally, and in the same context, Ayman El-Baz,
et al.[2], published in 2013 an extensive review paper with 46 pages on "Computer-Aided
Diagnosis Systems for Lung Cancer: Challenges and Methodologies" using 364 references
spanning 33 years of research. This review discussed extensively the subject of CAD systems for
lung cancer using these references, through the four main steps of processing in these systems:
segmentation of the lung fields(regions), detection of nodules inside the lung fields, segmentation
of the detected nodules, and diagnosis of the nodules as benign or malignant. It addressed several
issues like methodologies, training, testing databases and validation of methods used, identified
the challenges researchers faced, and the strength and drawbacks of the existing approaches.
Concerning systems for detection and diagnosis of lung cancer nodules, all the review papers we
mentioned above called for further improvements in the performance of the available systems by
improving the techniques they use. Specifically, reference [2] specified accurate segmentation of
lunge fields (lungs from chest image) and detecting lung nodules (segmenting and detecting
nodules in lung images) as challenges for further investigation in this field.
In this paper, we segment lungs from chest image accurately using what we called "lung mask"
produced using 2D Connected Components Labeling 2D-CCL followed by some morphological
operations, and a new method to segment lung nodules from extracted lungs image efficiently
using 2D-CCL to find regions that are probable to be nodules, then, by extracting shape , texture
and density features, necessary training is made, and the nodules are detected and extracted from
test data. Finally, a performance consistency-check is made by testing images of healthy persons
to show that the program detects no nodules in this cases.
2. RELATED WORKS
Different techniques were used by researchers to extract nodules from 2D and 3D, CT images.
Using two dimensional CT images, Kaur R., et al. [7] used PCA (Principal Component analysis)
to extract nodules from lung cancer CT images, and Miwa T., et al.[8] used morphological N-
Quoit Filter to automatically extract nodules based on shape and gray level information. Homma
N., et al. [9] used Gabor filter and the difference of pixel values along the object axis to detect
nodules, and Gomathi M., et al. [10] used FPCM and extreme learning machine for the same
purpose. Those were some sample references from the period (2002-2013). Recently, S. Makaju
et al. [11], used in 2018 watershed technique for segmentation and some shape and density
features to detect nodules., S. Wang et. al. [12] used in (2020) Residual Neural Networks and N.
Khehrahet. al [13] used in (2020) the histogram and some morphological operators to extract the

27
lung, and a threshold based technique to select candidate nodules. The works mentioned above on
detecting lung nodules from 2D chest CT images for Computer Aided diagnosis are naturally not
exhaustive but give a good idea about the diversity of techniques and methods used.
Ammar M. et al. [14], used the CCL technique to extract liver area from the complicated 2D
abdominal CT image to be used for diagnosis of liver cancer. Based on this experience, the
authors explored the possibility of using the same technique for the detection and extraction of
lung cancer nodules from 2D chest CT images, and presented the encouraging results they
obtained in this paper.
3. USED DATA
Images from CT scans of lungs of 11 persons were provided by Alsham Imaging Center and 102
others by Tishreen Hospital. The images in each scan are about 80, and the thickness of each slice
is 2 mm. The specialist selected one image from each scan to be used in this study. We divided
the 113 images into 2 groups: (1) lungs of 98 cancer patients containing nodules, and (2) lungs of
the remaining 15 persons with no nodules (from healthy persons). Both groups were used for
lung area extraction from the CT image. For nodule detection, we used the first group for
developing and testing the algorithm, and used the second group to check the consistency of the
algorithm performance, since the algorithm that detects nodules in the lungs of cancer patients
must detect "no nodules" in the images of the lungs of healthy persons. We transformed all the
images from DICOM format to JPG format for processing in MATLAB environment. Fig. 1
shows an example from the CT slices of a healthy person, and three other images from CT slices
of cancer patients lungs with different number of cancer nodules in each slice (4,8,1, respectively,
in raster order).
A lung CT image of a healthy person A lung CT image of a cancer patient (4
nodules)
534x512 703x512
A lung CT image of a cancer patient (8
nodules)
A lung CT image of a cancer patient (1
nodule)
548x512 657x512
Fig. 1. A CT image of a healthy person (upper left), a CT image of a lung cancer patient with 4 nodules
(upper right), a CT image with 8 nodules (lower left), and a CT image with 1 nodule (lower right). Nodules
are marked by circles by specialized doctor.

28
4. NODULES AUTOMATIC EXTRACTION METHOD
As can be seen in Fig. 1, the original lungs CT image is a complicated content one because it
contains, as well as the lungs area, the name of the medical center, the name of the patient, the
date, and several other types of information and shapes to help the doctor in diagnosis and
archiving. Besides, the lung nodes and the nodules are rather similar in shape, and their gray
levels are similar to those of the surrounding region. This situation makes direct extraction of the
lungs with their pictorial content from the original image (by thresholding, for example)
impossible. Therefore, the general method we use to extract the nodules consists of two main
stages, as shown in Fig. 2. Of course, each stage consists of several steps. In the first stage, the
lungs area is extracted from the original CT image, and in the second one, the nodules in the
lungs area are detected and extracted. If no nodules found in the lungs image, it is marked as
"healthy lungs". The output image contains the extracted nodules that can be used later in
diagnosis research.
Fig. 2. The two main stages of the general nodules extraction method with results.
4.1. Lungs Area Extraction Stage
The automatic extraction of lungs area from the chest CT image is a complicated process. It is
rich of details that we should consider in order to extract the lungs accurately from different CT
Extracted Nodules Image

29
shapes and sizes. We will explain this stage below in details due to its importance because
accurate extraction of lungs area affects greatly the correct detection and extraction of nodules in
the next stage. In this stage, we extract the lungs area from the original chest CT image through
two essential processes. In the first one, the known "lungs mask" is extracted through several
steps. Then the mask is multiplied by the original image to extract the lungs area from the chest
image with the original gray levels preserved. The block diagram in Fig. 3 summarizes this stage,
which will be explained below:

Fig. 3 A block diagram summarizes the lungs area extraction process.
Input chest CT image
Automatic thresholding
using Otsu's method.
Removing printed
characters.
Forming lungs mask.
Extracting lungs area with
the original gray levels.
Output image for
nodules extraction
Connected components
labeling and selecting the
largest component.
Removing organs that may
appear near the lungs.

30
The lungs area extraction process explained briefly in the block diagram in Fig. 3 is detailed
below:
1– Thresholding the input original image automatically using Outsu's method [15] to select the
threshold, then all pixels values in the image below this threshold are set to zero "0", and the rest
ones are set to "1", resulting in a binary image. Fig. 4(b) shows the result of thresholding the
original image shown in Fig. 4(a).
2- Labeling the connected components in the thresholded image to give each distinct object
(connected component) a unique label that enables us, in principle, to compute all possible kinds
of features and use them appropriately to extract lungs area or to distinguish the nodules.
3– Finding the largest component which is the closed region surrounding the lungs, by selecting
the component that has the maximum area measured by number of pixels in the thresholded
image, Fig. 4(c) shows the largest component obtained from Fig. 4(b).
4– Applying a "closing" process to the complement of Fig. 4(c) using a circular "structuring
element" with a diameter (D=8) to remove any remaining printed characters or tiny objects
attached to the largest component and to close small holes. The result of this process is shown in
Fig. 4 (d).
5 – Applying a hole-filling process to the complement of Fig. 4 (d) to get the complete inner area
of the largest component, shown in Fig. 4 (e).
6- Removing any organs may remain near the lung: This is done by applying an "opening"
process followed by a "closing" one using a circular "structuring element" with a diameter
(D=10). This case does not appear in some CT scans. Fig.4 (g) shows an organ removed from
Fig. 4(f) by this step. Note that we use here a different image to show this case, which does not
appear in all CT images.
7- Multiplying Fig. 4(d) by Fig. 4(e) to get what we called Lung Mask shown in Fig. 4 (h).
8- Finally, multiplying the Lung Mask by the original image in Fig. 4(a) to extract the lungs area
with original gray levels preserved, as shown in Fig. 4 (i).

31
Fig. 4. Sample results of the eight steps lungs extraction process.
The method explained above extracted successfully lungs areas from CT images of different
general shapes and complications, as shown in Fig. 5. These results give an idea of the flexibility
and effectiveness of the proposed method.
Original CT image Extracted lungs area Original CT image Extracted lungs area

32
Fig. 5. Sample results of extracting lungs areas from images of differentgeneral shapes, sizes, and
complication, with very good accuracy.
4.2. Extraction of Nodules from Lungs Image
After having the lungs area extracted, we extract nodules through three main stages: (1) Preparing
the lungs area image for feature extraction, (2) Extracting features used for classification, and (3)
Classifying the regions remaining in the prepared image into "nodules", and "not nodules".
Regions classified as "nodules" remain in the resultant image and the others are removed, as
shown in Fig. 2. We explain the three stages in the following in necessary details.
4.2.1. Preparing the Lung Area for Feature Extraction
Preparing the image of lungs area for feature extraction is done in three steps, as shown in Fig. 6.
These steps are:
(1) Binarizing the image using Outs's method to select the threshold in the same way used in
step 1 of lungs area extraction process explained in section 4.1.
(2) Labeling connected components in the lungs area binary image.
(3) Removing small components with areas less than 15 pixels, since as our investigation of all
images showed that the regions of such areas are not nodules. Removing these small
components will save some of the computation time needed for feature extraction and
classification. Fig. 7 shows a sample result with the remaining CCs.

33
:
Fig. 6. Preparing the lungs image used for feature extraction.
Fig. 7. An extracted lungs image (left), binarized image (middle), remaining CCs after small
components removing (right).
4.2.2. Feature Extraction
Specialized doctors diagnosed the nodules used in this study. We found by visual interactive
investigation of our data using MATLAP programming facilities that, in general, the nodules
have a rather circular shape with almost specific distribution of gray levels, Fig. 8 shows two
examples of the interactive examination we made. Therefore, we extracted two groups of
features: shape features group and density features group. We will explain the two groups in the
following two subsections.

34
Fig. 8. Two examples (gray level measuring (left) and threshold selection(right)) of the response of the
visual interactive image examination using MATLAP facilities we used.
4.2.2.1. Shape features used in the study
After labeling the remaining components in the lung image, we extracted the following shape
features for every component:
Fs1:𝑓𝑜𝑟𝑚 𝑓𝑎𝑐𝑡𝑜𝑟 =
4𝜋∗𝐴𝑟𝑒𝑎
𝑃𝑒𝑟𝑖𝑚𝑒𝑡𝑒𝑟
;Fs2: 𝑟𝑜𝑢𝑛𝑑𝑛𝑒𝑠𝑠 =
4∗𝐴𝑟𝑒𝑎
𝜋∗𝑀𝑎𝑥𝐷𝑖𝑎𝑚𝑒𝑡𝑒𝑟2
Fs3:solidity =
𝐴𝑟𝑒𝑎
𝐶𝑜𝑛𝑣𝑒𝑥𝐴𝑟𝑒𝑎
; Fs4:extent =
𝑇𝑜𝑡𝑎𝑙𝐴𝑟𝑒𝑎
𝐴𝑟𝑒𝑎 𝐵𝑜𝑢𝑛𝑑𝑖𝑛𝑔 𝑅𝑒𝑐𝑡𝑎𝑛𝑔𝑙𝑒
Fs5:compactness =
√(4∗𝐴𝑟𝑒𝑎)/𝜋
𝑀𝑎𝑥𝐷𝑖𝑎𝑚𝑒𝑡𝑒𝑟
; Fs6:aspect ratio =
𝑀𝑎𝑥𝐷𝑖𝑎𝑚𝑒𝑡𝑒𝑟
𝑀𝑖𝑛𝐷𝑖𝑎𝑚𝑒𝑡𝑒𝑟
Where Fs stands for: shape feature.
4.2.2.2. Density features
Density features are those related to gray levels of image components. Several density features
can be extracted from a gray image [16]. We extracted the following ones:
Fd1: 3rd
normalized moment: 𝜇2 = 255 ∗ ∑ (𝑧𝑖 − 𝑚)2
𝑝(𝑧𝑖)
255
𝑖=0
Fd2: 4th
normalized moment: 𝜇3 = 255 ∗ ∑ (𝑧𝑖 − 𝑚)3
𝑝(𝑧𝑖)
255
𝑖=0
Fd3: standard deviation: σ = √𝜇2
Fd4: Smoothness measure: 𝑅 = 1 −
1
1−𝜎2
Fd5: similarity measure: 𝑈 = ∑ 𝑝2
(𝑧𝑖)
255
𝑖=0
Fd6: entropy: 𝑒 = − ∑ 𝑝(𝑧𝑖)
255
𝑖=0 ∗ 𝑙𝑜𝑔2{𝑝(𝑧𝑖)}
Where: 𝑚 = ∑ 𝑧𝑖 ∗ 𝑝(𝑧𝑖)
255
𝑖=0 , Fd: density feature, m: average of gray levels; zi: value of the pixel
at gray level (i); p(zi):value of the histogram at gray level (i).
4.2.3. Detection of nodules from extracted lungs with remaining CCs
The image of the extracted lungs area with the remaining connected components CCs, like that in
Fig. 7, contains nodules and not-nodules CCs. Therefore, detecting nodules in this situation

35
belongs to the Pattern Recognition standard two-classes problem, in which a test sample must be
classified whether to belong to the first class (nodule, here), or to the second class (not-nodule).
This situation is exactly similar to signature verification problem in which the test signature must
be judged whether to be genuine or a forgery. Ammar M. developed a reliable system for
signature verification that gives its decision based on a threshold on a Weighted Euclidian
Distance Measure computed from suitable features [17]. This system is US patented [18], and
still working commercially in hundreds of US banks since about two decades examining about 3
million checks every day. Based on that, this approach was expected to work well in nodule
detection. We detected nodules-components by computing the Weighted Euclidean Distance
(WED) of each CC in the image using shape and density features explained in the previous two
subsections, then using a global threshold obtained from the training group, we classified each
CC with a WED less than this global threshold as a "nodule", because the low WED means that
the CC resembles the nodules more than the arbitrary CCs. CCs that have a WED equal or larger
than the threshold is considered as not-nodule.
In summary, the classification process is done as follows:
(1) Labeling the components (regions) remained in each image after small components
removal.
(2) Extracting all features mentioned above (shape and density ones) for every component
labeled in the image.
(3) Computing the WED measure for every component which is the Weighted Euclidean
Distance to the known cluster learned from group (1).
(4) Using a global threshold applied to the distance measures of all labeled components, if
the distance is less than the threshold, it is judged to be a "nodule", otherwise, it is judged
to be "not- nodule", and removed from the image.
(5) At the end of this process, the image will contain only the components judged as
"nodule", like those appearing in Fig. 2.
4.2.3.1. Types of decisions in this classification process and their impact on the results
There are three types of decision in the classification process mentioned above:
1 – True Positive (TP): a component judged to be a "nodule", and at the same time, it is
diagnosed by the doctor as "nodule".
2 – False Positive (FP): a component judged to be a "nodule", while it is not (not diagnosed by
the doctor).
3 – False Negative (FN): a component judged to be "not a nodule" and removed from the image,
while it is a nodule and diagnosed by the doctor.
TP is the desired decision because it is the correct one.
FP is not a correct one but it is not dangerous because it gives an alarm that there is a "nodule",
but in fact there is no nodule. The negative effect of this decision is that the doctor has to
examine the slice image for the availability of a nodule.
FN is a wrong decision because it overlooks an actual nodule available in the image. This can be
dangerous if the image contains only the nodule overlooked and removed from the image, and
this nodule is malignant . If there are more than one nodule and only one of them is detected, then
this case will not be dangerous because the specialist will examine this slice.

36
Keeping the above considerations about decisions and their impact on the results in mind, we will
discuss some representative experimental results we obtained very closely (at the individual CC
level). In this way, we can give the reader an honest and informative view of our results and their
analysis.
For view convenience, we will present detailed results using shape features only. Table 1 shows
feature values for a test image containing 4 remained CC including one diagnosed as nodule, as
well as the distance measure and the doctor diagnosis. "T" in the last column means that the
connected component number 3 is diagnosed as "nodule" by the doctor. The WED measure of
this CC is appreciably low compared with the other 3 ones which means that it can be easily
separated from them by a threshold, and consequently making a TP decision.
Table 1: Shape feature values, WED measure, and doctor diagnosis of connected components of an image
in the test data with 4 remaining components after small ones removal.
CC
#
Form
Factor
Roundne
ss
Solidity Extent
Compactne
ss
Aspect
ratio
Distanc
e
Doc.
Diag.
1 12.0391719
1001.605
65
0.83333
3
0.53571
4
31.6481540
2
2.1883096
97
1.53900
6
2 16.5141083
432.9014
45
1 0.9375
20.8062837
9
1.1365708
57
1.43579
5
3 31.7389353
23175.24
54
0.88695
7
0.65384
6
152.234179
3
1.2810900
21
0.20186
3
T
4 17.431828
7495.769
87
0.82
0.41414
1
86.5781142
5
2.4833007
8
1.30687
2
4.2.3.2. A Discussion inside the Classification Process
We will now proceed to more complicated case in which there are 40 remaining CCs and 5
diagnosed nodules (number: 12, 19, 34, 39, 40 in Table 2). Intuitively, the separation process
between diagnosed nodules and the other remaining CCs will be more difficult here. Table 2
shows the same information as in Table 1, just explained above, but for another image, naturally.
If we consider this table, and consequently its related image alone, we find that a threshold value
TH=1.39 on the WED measure will give the following result:
TP=100%, FP=6, FN=0.
This is a very excellent result since all diagnosed nodules are correctly detected, with 6 FPs and
no FNs. False positives in this case do not lay any new burden on the specialist because he must
examine this slice any way.
If we have to set one threshold for both images represented in Table 1 and Table2, we find that
this threshold (TH=1.39) works well but we will have one more FP from Table 1. The result in
this case becomes:
TP=100%, FP=7, FN=0.
Now if we want to minimize FPs, we may make TH=1.27. This will make the results using the
two slices (two persons) containing 6 diagnosed nodules as follows: TP=83%, FP=5, FN=1, that
is because nodule number 4 in Table 1 and number (25, 31, 32) in Table 2, will not be FP. On the
other hand, we have in this case 1 FN (number 12 in Table 2). However, this FN is not actually a

37
problem because 4 nodules are correctly detected and consequently the doctor must examine this
slice any way.
Moreover, if we wish to consider more slices (persons) using the same threshold (TH) for general
decision, we recognize that we must make a compromise between TP, FP, and FN, according to
some considerations we set.
For the final result we reported below in Table 3, we selected (automatically) the threshold on the
WED measure computed using shape features, and both of shape and density features, for all
connected components remained in images after removing small ones, so that we got the best
result.
Table 2: Shape feature values, distance measure and doctor diagnosis of connected componentsof another
image in the test data with 40 remaining components after small ones removal.
CC
#
form factor roundness Solidity Extent
Compactne
ss
Aspect
ratio
Distan
ce
Doc.
Diag.
1 10.4249357 3446.79225 0.7
0.2916
67
58.7093880
5
3.2595105
26
2.0301
08
2 17.4637494 119186.896 0.635714
0.3178
57
345.234552
1
7.0070551
13
6.6961
63
3 12.9455689 68606.4649 0.521368
0.2606
84
261.928358
3
7.1973141
77
5.2894
46
4 13.8022678 977.054729 0.9375 0.625
31.2578746
7
2.4682796
53
1.6177
06
5 10.2907557 9158.48268 0.6
0.2596
15
95.6999617
6
4.3508633
56
2.6314
02
6 12.5218755 21887.498 0.625
0.3508
77
147.944239
4
5.3074196
23
3.2177
01
7 19.9771266 80344.6825 0.807339
0.2857
14
283.451375
9
5.8931146
33
4.7493
8 18.4375056 6146.43619 0.904762
0.5757
58
78.3992104
1
2.4200665
16
1.2619
81
9 31.6384011 17473888.8 0.251653
0.1476
41
4180.17808
7
2.3879955
92
834.81
04
10 22.0959009 12503.9979 0.929825
0.5520
83
111.821276
6
2.5699683
4
1.1242
57
11 14.4832858 1526.04864 0.904762
0.5428
57
39.0646724
8
2.3723964
84
1.5348
69
12 13.8022678 631.237785 0.9375 0.6
25.1244459
6
1.4446228
11
1.3823
21
T
13 14.259656 5131.94222 0.878788
0.3580
25
71.6375754
9
3.3559262
03
1.9553
88
14 24.5138915 1021218.48 0.501157
0.3200
3
1010.55355
2
1.4175012
82
48.283
74
15 15.3966423 709.030683 1
0.8333
33
26.6276300
8
1.8691310
01
1.4424
22
16 11.8582659 3916.56401 0.846154
0.3666
67
62.5824576
9
4.4242579
92
2.7194
78
17 9.74569055 7612.43347 0.621622
0.5476
19
87.2492605
8
5.9885003
93
3.7869
39
18 10.2644282 14561.2714 0.535714 0.375
120.670093
3
5.4132930
88
3.3291
7

38
19 18.1783462 1143.61606 0.958333
0.6388
89
33.8173928
5
1.2420038
01
1.2674
55
T
20 12.6698853 2963.42367 0.875 0.7
54.4373370
5
3.5924716
41
2.1435
3
21 19.0638325 4343.9594 0.897436
0.6481
48
65.9087201
8
1.9122354
77
1.1141
46
22 14.2623434 7875.69015 0.815789
0.4305
56
88.7450852
2
4.4705205
41
2.6397
14
23 9.77406321 8358.06802 0.520833
0.3571
43
91.4224699
9
4.1934510
23
2.5212
88
24 37.7743416 2065085.55 0.460245
0.3284
23
1437.04055
3
1.7471527
27
98.216
22
25 18.3437668 1030.94052 0.956522
0.7333
33
32.1082624
3
1.2493024
61
1.2806
73
26 20.7676771 145079.296 0.582938
0.2971
01
380.892762
3.1947942
7
6.4433
88
27 16.2469754 3922.51568 0.875 0.7 62.6299903
2.8880397
32
1.6364
73
28 15.6814289 1901.00829 1 0.875
43.6005538
2
2.4986703
81
1.5764
98
29 13.4273865 14015.7852 0.698113
0.2569
44
118.388281
3
4.4521813
94
2.6168
38
30 22.6416841 61562.2386 0.679245
0.4060
15
248.117388
7
2.2825433
01
2.3289
33
31 17.4495279 610.07542 1 0.85
24.6997048
6
1.2631643
45
1.3655
54
32 23.3841595 12049.1814 0.916667
0.5555
56
109.768763
4
2.3675117
03
0.9888
34
33 17.3240175 26716.5399 0.707865 0.375
163.451949
9
3.2683634
46
1.7022
42
34 25.4311313 7373.25296 0.835616 0.61
85.8676479
1
1.1157040
69
0.7709
38
T
35 11.7170336 180720.806 0.57971 0.4
425.112697
4
11.341071
17
11.161
44
36 26.8874838 16670.4559 0.985294
0.7362
64
129.114119
9
2.2527588
63
0.7910
35
37 20.657524 4675.95165 0.972973 0.72
68.3809304
2
2.1643765
77
1.1829
8
38 17.3879834 13401.8703 0.862745
0.3666
67
115.766447
3
3.8511189
66
2.1107
13
39 23.1133404 2157.21731 1
0.8095
24
46.4458534
8
1.1272559
08
1.1613
56
T
40 30.6282678 9547.1584 0.985507 0.85
97.7095614
4
1.2569720
84
0.6689
26
T
4.2.3.3. Training
The training to obtain the global threshold is done on 30 images among the 98 ones containing
nodules diagnosed by the specialized doctor. If we refer to Table 1 again, it is obvious that the
CC number 3 is easily separable by a threshold from the other CCs due to the considerable
difference in the WED value, and can be detected as a nodule. This is an example, but the
complete training was done using all shape features, and both of shape and density features in 30
images from the 98 images containing nodules used in this study. It is worth noting that the

39
number of remaining CCs in an image reaches 41 in some images. We have already shown the
results of 40 remaining CCs in Table 2.
5. RESULTS AND DISCUSSION
After training on the 30 images using shape features once, and both shape and density features
once again to obtain the global threshold that separates the nodules from the other CCs in each
case, we obtained the results shown in Table 3, where: TP is "True Positive", "FP" is "False
Positive", "FN" is False Negative, and the "Sensitivity" is computed by this equation: Sensitivity
= TP/(TP+FN).
Table 3. Results of the nodules detection using shape and density features on test cases of 68 images.
Feature kind Sensitivity% FPs
(per case)
No. CCs
No. of Nodules
Diagnosed by
doctor
Shape features 95.1 2.1 1836 173
Shape +Density features 97.2 1.98 1836 173
We can conclude from Table 3 that: the average diagnosed nodules is about 2.54 per image, and
the average number of remaining CCs in an image before detection and extraction is 27.
Using the density features with shape ones improves the sensitivity by about 2%, and the
improvement in FPs reaches about 6%.
5.1. Performance consistency check
As a consistency check of the performance of the nodule detection and extraction algorithm, we
tested the lung images of the 15 healthy persons as a consistency-test-group. The result was
excellent where 9 false positives (FPs) appeared in 8 images (4% of CCs, as an average). Fig. 9
shows the result for a test image containing seven nodules where all detected with one FP, and
the second is for a healthy person where no nodules are detected, and consequently will be
classified as "healthy lung" .
Fig. 9. The first column is the input image, the second column is the extracted lungs image, and the third
column is the result of detection and extraction of nodules. In the first row: TPs=7, and FPs=1. In the 2nd
row: TPs=0, FPs=0, healthy person (perfect result).

40
5.2. Comparison with some other works
We compared our method with 9 works conducted in the period 2010-2020 as shown in Table 4
below. We can see in this table that the introduced method gave a general performance
comparable with the best results reported, with little bit higher sensitivity, and the FPs rate is
among the low ones.
Table 4. Performance comparison between other works and the introduced method.
Authors year Sensitivity% FPs per case No. nodules
Liu [19] 2010 97 4.3 32
Cascio [20] 2012 97 6.1 148
Orozco [21] 2012 96.15 2 50
Teramoto [22] 2013 80 4.2 103
Shao [23] 2012 89.47 11.9 44
Bergtholdt [24] 2016 85.90 -
Wu [25] 2017 79.23 -
Saien [26] 2018 83.98 0.02
Khehrah [12] 2020 93.75 0.13
Monif(our work) 2021 97.2 2.1 173
6. FURTHER INVESTIGATION AND FUTURE WORK
In an attempt to find some way to improve the performance of the method further, we examined
the images containing nodules and compared them with the numerical results. We found that a
considerable part of wrong decisions is related to what we can call Nodules Attached to the
Shield (NAS), where NAS is explained as follows: In Fig. 10 below, we can see in the upper
right image two nodules diagnosed by the specialized doctor (surrounded by circles). Those two
nodules are attached to the shield (white area surrounding the lungs) and called "NAS".
A chest CT scan slice containing two nodules
attached to the shield (NAS), as diagnosed by
the specialized doctor .
Extracted lungs area
The two extracted nodules, with one of them
wrongly extracted as FP (upper right), and one
of those diagnosed by doctor (bottom wright
one) was not extracted (FN) .
Two nodules Demarcated on lungs area by
the program ( our method).
Fig. 10. Two nodules extracted by the method in the right bottom image, one of them (upper one) is FP.
And one of the two nodules diagnosed by the doctor was not detected (FN).

41
We plan to work on improving the performance in two directions:
1 – Improving the segmentation of the NAS,
2 – Improving the correct classification rate ( sensitivity ) by using the feature selection
technique developed by Ammar M. [17].
7. CONCLUSIONS
We have introduced a method for the automatic detection and extraction of lungs cancer nodules
from CT images using connected component labeling (CCL) and weighted Euclidian distance
measure based classification. The obtained results have shown clearly the high performance of
the method in both extraction of lungs area from the CT image and in the correct detection of the
cancer nodules. This high performance was also supported by the consistency check we made
which reveals the ability of the method to detect the healthy image at the same time.
The experiments conducted and their results analysis presented in this paper enable us to
conclude that using CCL technique with appropriate sequence and parameters of some
morphological operations may lead to a high performance approach for extracting lungs areas
from complex chest CT image with a wide variety in shapes and sizes. We can conclude also that
using CCL technique with the high flexibility it offers in manipulating CCs in the extracted lungs
areas, with the usage of WEDM and a threshold based classification, may provide an efficient
method for detecting and extracting nodules to be used later for diagnosis. We found also that
using density features with shape ones improves to some extent the performance. Our further
investigation have also shown that improving the segmentation of NAS may improve the final
results. We have set this point as a future work.
We hope that we have introduced a positive effort in the general direction of the research for
building an actual automatic lung cancer detection and diagnosis systems.
ACKNOWLEDGMENTS
The authors wish to thank Alsham Medical Imaging Center and Tishreen Hospital for providing
the lung images data. We appreciate also the support provided by Al Andalus University for
Medical Sciences and its hospital.
REFERENCES
[1] WHO, Latest world cancer statistics, (2013), The International Agency for Research on Cancer
Publications, Geneva: World Health Organization.
[2] Ayman El-Baz, et al.,(2013) Computer-Aided Diagnosis Systems for Lung Cancer: Challenges and
Methodologies, International Journal of Biomedical Imaging, Volume 2013, Article ID 942353.
[3] Heang-Ping Chan et al, (2008) Computer-Aided Diagnosis of Lung Cancer and Pulmonary Embolism
in Computed Tomography — A Review, Acad. Radiology, 15(5): 535–555.
[4] Sprindzuk M.V., et al., (2010) Lung cancer differential diagnosis based on the computer assisted
radiology: The state of the art, Pol J Radiology, 75(1): 67–80.
[5] Firmino et al., (2014) Computer-aided detection system for lung cancer in computed tomography
scans: Review and future prospects, BioMedical Engineering, 13:41.
[6] Bhavanishankar K. et al., (2015) 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.

42
[7] Kaur R., and Ada S. (2013), "Feature Extraction and Principal Component Analysis for Lung Cancer
Detection in CT scan Images", International Journal of Advanced Research in Computer Science and
Software Engineering, Vol. 3, Issue 3, pp.187-190.
[8] Miwa T., Kako J., Yamamoto S., Matsumoto M., Tateno Y., Iinuma T., and Matsumoto T. (2002),
"Automatic Detection of Lung Cancers in Chest CT Images by the Variable N-Quoit Filter", Systems
and Computers in Japan, Vol. 33, No. 1, pp.53-63.
[9] Homma N., Takei K. and Ishibashi T. (2008), "Combinatorial Effect of Various Features Extraction
on Computer Aided Detection of Pulmonary Nodules in X-ray CT Images", INFORMATION
SCIENCE & APPLICATIONS, Issue 7, Vol. 5, pp.1127-1136.
10] Gomathi M., and Thangaraj P. (2010), "A Computer Aided Diagnosis System For Detection Of Lung
Cancer Nodules Using Extreme Learning Machine", International Journal of Engineering Science and
Technology, Vol. 2, pp. 5770-5779.
[11] Wang S., Dong L., Wang X., Xin Wang, Open Med. (2020) Classification of pathological types of
lung cancer from CT images by deep residual neural networks with transfer learning strategy.
[12] Khehrah N., Farid M. S., Bilal S. and Khan M. H., (2020) Lung Nodule Detection in CT Images
Using Statistical and Shape-Based Features, J. Imaging, 6, 6 (14 pages).
[13] Makaju S., Prasad P.W.C., Alsadoon A., Singh A. K. and Elchouemi A., (2018) Lung Cancer
Detection using CT Scan Images, Procedia Computer Science 125- 107–114.
[14] Ammar M., et al., (2011) Using Image Processing Techniques for Automatic Extraction of Liver
Suspicious Regions from X-Ray Computed Tomography Images, Tishreen University Journal for
Research and Scientific Studies - Engineering Sciences Series, Vol. (33) No. (3), pp.(217-235).
[15] Nobuyuki Otsu (1979). "A threshold selection method from gray-level histograms". IEEE Trans.
Sys., Man., Cyber. 9(1): 62–66.
[16] Gonzalez, R., Wood, E. and Eddins, S. (2009) "Digital Image Processing Using MATLAB", Second
Edition, Gatesmark Publishing.
[17] Ammar M., Raising the Performance of Automatic Signature Verification Over that Obtainable by
Using the Best Feature Set, (2011) International Journal of Pattern Recognition and Artificial
Intelligence. 25-2, PP 183-206.
[18] Ammar M., Method and apparatus for verification of signatures, United States Patent: No. 6424728 ,
07/23/2002, U.S.A.
[19] Liu Y, Yang J, Zhao D, Liu J (2010) A method of pulmonary nodule detection utilizing multiple
support vector machines. In Computer Application and System Modeling (ICCASM), International
Conference On, vol. 10. Taiyuan; 10–11810121.
[20] Cascio D, Magro R, Fauci F, Iacomi M, Raso G (2012)Automatic detection of lung nodules in CT
datasets based on stable 3d mass-spring models. ComputBiol Med, 42(11):1098–1109.
[21] Orozco HM, Osiris Vergara Villegas O, Maynez LO, Sanchez VGC, de Jesus Ochoa Dominguez H
(2012) Lung nodule classification in frequency domain using support vector machines. In Information
Science, Signal Processing and Their Applications (ISSPA), 11th International Conference On.
Montreal, QC; 870–875.
[22] Teramoto A, Fujita H (2013) Fast lung nodule detection in chest CT images using cylindrical nodule-
enhancement filter. Int J Comput Assist RadiolSurg, 8(2):193–205.
[23] Shao H, Cao L, Liu Y (2012) A detection approach for solitary pulmonary nodules based on CT
images. In Computer Science and Network Technology (ICCSNT), 2nd International Conference On.
Changchun; 1253–1257.
[24] Bergtholdt, M.; Wiemker, R.; Klinder, T. (2016) Pulmonary nodule detection using a cascaded SVM
classifier. InProceedings of the Medical Imaging: Computer-Aided Diagnosis, San Diego, CA, USA,
27 February–3March 2016; Volume 9785, pp. 268–278.
[25] Wu, P.; Xia, K.; Yu, H. (2016) Correlation coefficient based supervised locally linear embedding for
pulmonary nodule recognition. Comput. Methods Programs Biomed. 136, 97–106.
[26] Saien, S.; Moghaddam, H.A.; Fathian, M. (2018) A unified methodology based on sparse field level
sets and boosting algorithms for false positives reduction in lung nodules detection. Int. J. Comput.
Assist. Radiol. Surg., 13, 397–409.

43
AUTHORS
Maan Ammar Ph. D. in Information Engineering, Nagoya University, Japan, 1989,
Professor at Al Andalus University for medical sciences, Biomedical engineering since
2014, Full professor at Applied Sciences University, Amman Jordan 2003, US patent of a
commercial system serving hundreds of US banks since 2002 "Method and apparatus for
verification of signatures", United States Patent: No. 6424728 , 07/23/2002, U.S.A.
Published many papers in image processing and pattern recognition fields. Served as
Head of biomedical engineering department–Damascus University for 8 years.
Muhammad Shamdeen Received B. SC in Biomedical Engineering in 2014 from
Damascus University. Now he is a Master student in the biomedical department at
Damascus University. Working in the field of medical image processing.
MazenKasedeh Received B. Sc in Biomedical Engineering in 2014 from Damascus
University. Now he is a Master student in the biomedical department at FMEE. He is
Interested in medical image processing.
Kenan Mansour MD Obstetrics and Gynaecology Specialist, Tishreen Hospital, 2011.
Medical Manager of Al Andalus University Hospital, 2019 to present. Medical Education
Master student, Syrian Virtual University, 2020. Interested in medical image analysis and
diagnosis.
Waad Ammar MD General Surgery Specialist, Tishreen Hospital, 2020. At present,
working at Al Andalus University Hospital. Medical Education Master student, Syrian
Virtual University, 2020. Interested in medical image analysis and diagnosis.

Using Distance Measure based Classification in Automatic Extraction of Lungs Cancer Nodules for Computer Aided Diagnosis

More Related Content

Similar to Using Distance Measure based Classification in Automatic Extraction of Lungs Cancer Nodules for Computer Aided Diagnosis (20)

Recently uploaded (20)

Using Distance Measure based Classification in Automatic Extraction of Lungs Cancer Nodules for Computer Aided Diagnosis