2018 22nd International Conference on System Theory, Control and Computing (ICSTCC)

Lung Cancer Diagnosis based on Ultrasound image

processing

Achim Cristian, Rusu-Both Roxana, Dulf Eva- Chira Romeo Ioan

Henrietta st
I Medical Clinic, “Iuliu Haieganu” University of Medicine
Automation Department and Pharmacy Cluj-Napoca
Technical University of Cluj-Napoca Cluj-Napoca, Romania
Cluj-Napoca, Romania romeochira@yahoo.com
achimcristian@yahoo.com, roxana.both@aut.utcluj.ro

Abstract— Cancer represents one of the leading causes of symptoms, even if some are nonspecific, which causes them to
mortality in this century. Patients suffering from lung cancer contact the doctor. Symptoms may be the consequence of the
(LC) have an average of 5 years life expectancy after diagnosis. primary tumor, which can cause local compression or can
This is due usually to late detection. The most common methods spread through metastasis to the rest of the body. In some
to detect lung cancer are computed tomography (CT) and cases, the primary tumor does not have a pulmonary location,
magnetic resonance imaging (MRI). These investigations are but it appears secondary, which is why the clinical picture can
invasive to the human body and are performed only based on a vary greatly from patient to patient. Cough is the most common
doctor’s recommendation, which comes usually after patients symptom of lung cancer. However, many smokers (those who
develop symptoms – hence the late detection of cancer. Nowadays
smoke for a long time) have chronic cough, so it is very
transthoracic ultrasonography (TUS) and US-guided biopsy have
gained a larger field in the management of patients with
important for a person with chronic cough to go to the doctor's
peripheral pulmonary nodules or masses. From the multiple control if they notice that the cough is changing or getting
ultrasonography (US) advantages which made it a well- worse.
established solution in the management of tumoral and non- There are two main types of lung cancer: small cell lung
tumoral abdominal pathology, few stand out: it’s non-invasive cancer and non-small cell lung cancer. The latter is the most
and has reduced costs. However, until now transthoracic US is common type and it can spread anywhere in the body through a
underused for lung cancer diagnosis because it can be hard to
process called metastasis. There are various methods for
interpret to determine an accurate diagnosis. In this paper we
diagnosing lung cancer and the most common used are:
aim to develop a software application for lung mass classification
based on ultrasound image processing. This could be an
computerized tomography (CT) scan, positron emission
important step in towards early detection of lung cancer, by tomography (PET) scan, magnetic resonance imaging (MRI)
introducing the transthoracic ultrasonography in the regular scan. In the case of PET scan [3, 4], it involves radioactive
annual check-up, improving the life expectancy of patients or tracers, but the exposure to radiation is minimal.
even complete recuperation. Ultrasonography (US) [3, 5] has been developed in order to
offer a lot of useful information about the tumors such as the
Keywords—lung cancer, ultrasonography, tumor detection,
tumor structure, vascularization, the stage of parietal invasion
image processing
and sometimes lymph node invasion. The US is less expensive
I. INTRODUCTION and comparatively safer technique than CT scans. More than
that, it is not recommended for the human body to perform a
Lung cancer (LC) represents a leading cause of cancer- CT, rather than by the doctor’s recommendation, because for a
related mortality worldwide. Recently published data estimates person it would take 3.3 years to get the same amount of
an increase of LC deaths worldwide from 1.6 million in 2012 background radiation that a CT delivers in less than a minute
to 3 million in 2035, [1]. It is the second most common cancer [3, 6]. Even considering the advantages of the US, it is not used
in both men and women due to the tobacco consumption. This for lung cancer detection. The main reason is that is very
is the main cause, but not the only one, because lung cancer is a difficult to interpret by lung specialists that are not used to
disease that can affect anyone. In some cases, previous lung perform transthoracic ultrasonography on patients. Plus, the
diseases (chronic bronchitis, emphysema, pneumonia, and lung is not visible on US if it is healthy and full of air.
tuberculosis) could have a role in the development of lung
cancer [2, 3].The assessment of patients with pulmonary focal Taking into consideration the multitude of advantages the
abnormalities continues to represent a major problem in LC US has to offer, we imagined an application which helps
diagnosis. doctors diagnose lung cancer. Hence, the main contribution of
this work consists in the development of a lung cancer (LC)
Unfortunately, the early stages of lung cancer may not diagnosis software application able to accurately and non-
show symptoms. As tumors grow in size, they can produce a invasively detect non-small cell lung cancers as part of a
variety of symptoms. There are patients who show certain research project. Until now, we developed a preliminary

978-1-5386-4444-7/18/$31.00 ©2018 IEEE 73

software application for intrathoracic tumor recognition of Based on the clinical experience of the physicians we were
ultrasonography imaging, addressed to peripheral LC able to identify the main features that we need to extract from
diagnosis, based on US principles of definition of LC the ultrasonography image in order to be able to help the
concerning echogenicity, contours, echo structure, interfaces diagnostic process or even give a probabilistic diagnostic.
with adjacent structures, and also on diagnostic criteria for Based on their importance for the diagnostic process the main
differential diagnosis (lung condensations, cysts, benign tumors features of the determined area of interest are: A) the
and pleural pathology). This would help physicians to identify irregularities (spiculation); B) the shape; C) the transition rate
patients with LC and choose the next appropriate method of of the contour pixel intensities; D) the number/area of the
investigation – lung percutaneous biopsy, bronchoscopy, echo- whiter spots inside the contour and E) the aspect ratio of the
endoscopy, surgical treatment or surveillance. contour shape. In this step, in the development of the
diagnostic software application was implemented a feature for
The applied method uses the edge-based active contour global analysis and local contour characteristics, which can act
algorithm used for segmentation, similar to Geodesic Active as support for the radiologist and highlight the locations of
Contour [7] and it is a very flexible and powerful method. It suspect areas so that the overall sensitivity is high. The
can segment many types of images, including some that are proposed method for implementation consists of 2 processing
very difficult to segment with other methods based on
steps, Fig.1. The first step is to preprocess the ultrasound image
thresholding or gradient. We will outline the steps in which the and extract the region of interest (ROI) from the whole image.
application runs, the algorithm for segmenting images, and The second step involves detecting the contour of the tumor in
examine performance on some US images. the region of interest in order to classify it. The last step is to
II. PROPOSED METHODOLOGY extract the previously presented key features.
Driven by the fact that there are reports in which the
greyscale histogram was used for breast tumors, we performed
a pilot study as part of a national research project, in order to
evaluate the possibility to build a computerized model to assist
Fig. 1. Lung cancer tumor detection algorithm
young doctors or ultrasound technicians in recognizing
peripheral lung tumor and differentiate them from benign At this stage, the edge-based active contour algorithm was
condensations. chosen for implementation and testing for separating the
The first stage of the pilot study consisted in the suspicious region from its background to be used further for
identification of potential cases, their partial enrollment in the tumor classification stage. Tumor contouring is an important
study, the collection of blood samples as well as the acquisition criterion in classifying them into cancers or non-cancers.
of ultrasound images. As result, we analyzed two populations
of patients – with peripheral lung cancers and with lung III. LUNG CANCER SOFTWARE APPLICATION
condensations. We performed transthoracic US, recording The acquisition stage of the ultrasound images (the first
different parameters – diameters, contour, echogenicity, stage in the lung cancer detection algorithm, Fig.1) was
distribution of the vascularization, invasion of the adjacent performed using a General Electric LOGIQ S6 in .bmp or .jpg
structures, etc. Images were recorded and transferred for format.
computerized analysis. The development of lung cancer
diagnostic software could improve the quality and accuracy of The ultrasound image preprocessing step consists in
the diagnosis, enhance the success of therapy by early detection detecting the region of interest. This is done through the user
of cancer, avoid unnecessary biopsies and reduce radiologist's application. The following processing steps will apply only to
interpretation time. the region of interest determined by the user (physician).

Hence, the second stage of the pilot study consisted in the The entire software application was created in Matlab,
elaboration of the diagnostic software specifications that using its native image processing capabilities and complex
accurately detects cancer cells without any confusion, in order functions on top of which we implemented our own algorithms
to achieve the previously presented goals. The main features for feature. The algorithms were conjoined through the visual
that allow differentiation between benign tumors and malignant composition and temporal behavior of a graphical user
tumors are related to the shape of the nodules, the echogenicity interface, in order to enhance the efficiency and ease of use for
of the different regions and the regularity of their contours. An the underlying logical design of the program.
ellipsoid form or the presence of two or three well-defined lobs After the application started, the user has to load an
is considered benign finding. On the other hand, the presence ultrasound image and define a relatively accurate mask by
of ramifications or a form that is higher than wider are selecting the region of interest, from which the contour
considered to be malignant findings. Moreover, when algorithm will find its exact real margins, Fig.2. When acoustic
examining the smoothness of the contour, if it has micro-lobes, shadow is present on the image, a signal void behind structures
angular or spicular edges, they are interpreted as malignant like bones that strongly absorb or reflect ultrasonic waves, the
findings as opposed to a rounded and well-defined outline. user also has the possibility to add one or more white lines to
Finally, hyperechogenicity and the presence of an echogenic the image in order to separate those irrelevant areas. Having a
thin capsule increase the likelihood of benignity, while high intensity, the line stands in contrast with its surroundings,
calcifications and hypoechogenicity are associated with acting like a barrier. However, they only appear for this sole
malignant nodules. purpose, and not for the followed algorithms.

74
 The similar to Geodesic edge-based active contour farthest away from the centroid point, is plainly the largest of
algorithm evolves the segmentation using an iterative process, all D distances. The second point, as well as the third, must be
performing the number of iterations given by the user, 200 situated as relatively opposite to this first one and to each other
being the default. The algorithm starts from the given mask, a as possible. Therefore, the second point has to be located on the
binary image that specifies the initial state of the active contour not closer than the sixth of the entire contour length, by
contour, and then grows outwards, expanding until it finds either side of the first point. Likewise, the third one must be at
object boundaries. In order to obtain faster and more accurate a length/6 distance of each of the other two points. This
segmentation results, an initial contour position that is close to algorithm assures the largest distance between the three corner
the desired object boundaries should be specified. points of the inscribed triangle, offering its largest area. The
circumscribed circle is then constructed exactly on these three
vital points, encapsulating the entire contour. If the difference
between this circle and the contour is larger than the triangle-
contour difference, the shape tends towards a more circular
one, or triangular otherwise.
C. Fade Algorithm
Another one of the five key features that has to be studied is
the differences of pixel intensities at the lower part of the
contour. This feature is also very important because a hard in
depth transition of the pixel intensities indicates the cancerous
nature of the lung lesion in opposition to a smooth in depth
transition. To obtain a coefficient in this sense, two different
brightness averages were computed from the pixels around the
Fig. 2. Diagnostic application users interface (In red is the defined starting
mask and in green is the obtained contour after the segmentation)
outline of the shape. One mean from the points found inside the
contour and the other for the ones outside, but only around its
The results obtained using the implemented key features lower outline, the top part not being of interest in this fade
extraction algorithms are displayed in the left lower part of the determination algorithm, Fig.3. The final result is obtained
detection system user interface. Having the exact uneven from the difference between these two means. If this
outline of the real region of interest, the centroid is easily coefficient is relatively small, meaning close to zero, the fade is
computed as well as the distances (D) between it and all the a slow one, barely visible, compared to the sudden intensity
contour pixels. For the five key factors that had to be analyzed change when the large result is obtained, the contour
in order to determine if the region of interest has stronger representing the border of a high contrast in brightness.
characteristics of a cancer or a non-cancer, the main FadeCoeff = (FadeCoeffExt/FadeCoeffInt-1)
algorithmic program was structured accordingly.
D(i) = Distance(ContourX(i),ContourY(i),CentreX,CentreY)
A. Contour Algorithm
The more non-superficial irregularities the contour has, the
more it indicates towards a tumorous nature. Therefore, once
the contour was determined, we developed an algorithm to
determine the smoothness of the lesion’s contour by computing
the differences between the peaks and valleys of the contour in
respect of the centroid. The output of the algorithm is a
coefficient obtained from the mean of all these differences. Fig. 3. Fade algorithm example
Perfect smoothness, meaning no differences between all the
supposed D fluctuations, would indicate not only the absence D. White Spots Algorithm
of peaks and valleys, but also the shape of the contour being a The conglomerates of higher intensity pixels inside the
circle of radius D. Consequently, the Contour Algorithm also contour represent the white spots. From a diagnostic point of
plays a part in determining the shape of the studied area. view the presence of whiter spots indicates the presence of
ContourCoeff = mean(abs(difference([peaks; valleys]))) healthier areas inside the tumor’s contour. The desired result is
one regarding their computed area compared to the whole
B. Shape Algorithm lesion contour surface. If the computed area of the brighter
Because the shape and the smoothness of the lesion contour pixels is high compared to the formation shape surface, it
are the most important features in distinguishing its nature, indicates the tumorous nature of the lesion. To find and obtain
another approach to the deduction of its shape is the the number of the brighter pixels, all points inside the contour
comparison of its surface to that of the circumscribed circle and are compared to the average intensity of that surface. The ones
the inscribed triangle (which can indicate pneumonia). To brighter than that threshold are added up and compared to the
determine the inscribed triangle, three points and their others through a ratio. In order to not count in also the brighter
coordinates are needed in such a way to obtain the largest pixels which are on the contour line, as part of the sudden fade,
possible area for the triangle. The first point, specifically the

75
the area studied is taken slightly smaller than that of the The second test scenario was to use the developed
shape’s outline. diagnostic system on new ultrasound images of both confirmed
cancers and other non-cancers to test if based on the
BrightPix = sum(Img(ROImg >= mean2(ROImg(shape==0)=0 > 0)))
determined nominal features coefficients range value we will
E. Aspect Ratio Algorithm obtain the confirmed diagnostic. The second test scenario was
The aspect ratio of the lesion contour is defined as the ratio performed on multiple ultrasound images, some results
between the largest horizontal length and the largest vertical examples being presented in Fig.5 and Fig.6.
length, specifically the anteroposterior and latero-lateral axes.
This feature was implemented to identify if the shape is a
triangular one with the farthest point away from the surface,
like in the case of pulmonary condensation. To obtain this
ratio, each pair of opposite pixels must be found and the
distance between them computed. For the horizontal line, the y
coordinates range only by incrementing and the x coordinates
are the ones of the contour points at that precise index, Fig.4.
From all the distances between these coordinates, the largest
one is divided by the vertical line, obtained analogous.
AspectRatio = abs(max(verticals(ii))/max(horizontals(ii)))

a.

b.

Fig. 4. Shape and aspect ratio algorithm example

IV. RESULTS AND DISCUSSION

Using the developed application, a series of tests were
performed to determine the accuracy and degree of detection of
the contour of the lung lesions using ultrasound images of c.
confirmed malignant tumors (cancers) and other non-cancers.
To this end, the first test scenario was to use over 30
ultrasound images of both confirmed malignant tumors and
other non-cancers to determine each individual feature
coefficients using the developed diagnostic system. Based on
the performed tests we determined the nominal range value for
the feature coefficients in both cases, table I.

TABLE I. FEATURE COEFF. NOMINAL RANGE VALUES

Confirmed
diagnostic Cancers Non-cancer
Feature
A. Controur < 14 > 14
B. Shape circular triangular
C. Fade >5 <5
D. White spots > 600 < 600 d.
E. Aspect ratio 0.5-1.5 ‫ ב‬0.5-1.5 Fig. 5. Key feature coefficients for necrotic tumor: a) contour coeff; b)
white spot coeff; c) fade coeff; d) shape and aspect ratio coeff.

76
 As it can be observed in Fig.5 the developed diagnostic
system is tested on a confirmed malignant tumor
ultrasonography. In Fig.5.a can be observed the determined
contour of the tumor is represented with green color. The
contour represented with red color line is the contour of the
region of interest defined by the user (physician). The straight
white line that can also be observed are also defined by the user
just to eliminate the non-important areas for the contour b.
detection algorithm. Fig.5.b shows the sequential results based
on which the white spots coefficient is computed. Fig.5.c.
actually presents the dimension of the area outside the tumor
contour which is taken into consideration for fade coefficient
computation. Lastly in Fig.5.d is presented the circumscribed
circle, the inscribed triangle for the determined tumor contour
and also at the bottom side is presented the alignment of D,
specifically all the distances from the centroid to each pixel on
the contour, used for finding the peaks and valleys of the
outline. Based on these results, on the bottom left side of the
diagnostic system main window, Fig.5.a, are displayed the
determined key feature coefficients. As it can be observed
based on the obtained results and the key feature nominal
coefficient range value, Table I, the tumor can be classified as c.
malignant as the confirmed diagnostic.
700
In Fig.6, the obtained results after testing the developed
diagnostic system on a confirmed pulmonary condensation are 600
presented. In Fig.6.a can be observed the determined contour of
the suspicious region of interest. As in the previous case the 500
contour of the region of interest defined by the user (physician)
is represented with the red contour. In a similar manner Fig.6.b 400
shows the sequential results based on which the white spots
coefficient is computed. As it can be observed in this case the 300

number of the brighter pixels is greater. In Fig.6.c. are

200
presented the obtained results by applying the fade algorithm
and in Fig.6.d are presented the obtained results by applying
100
the shape and aspect ratio algorithms. As it can be observed in
the diagnostic system main window, Fig.6.a, in this case the 0
diagnostic system indicated shape is a triangular one. 0 100 200 300 400 500 600 700 800 900

Considering this results, together with the other key features

coefficients the suspicious region of interest can be classified d.
as benign which corresponds with the confirmed diagnostic.
Fig. 6. Key feature coefficients for pulmonary condensation: a) contour
coeff; b) white spot coeff; c) fade coeff; d) shape and aspect ratio coeff.

Based on the results presented so far we can state that the

developed lung cancer diagnostic system proved to be effective
by showing promising results in the test scenarios. The next
step in its development is to determine a mathematical model
based on the key feature coefficients in order to be able to
provide a possible diagnostic. A possible approach is to use
machine learning, neural networks in order to provide a
possible diagnostic. A minus of this approach is the need of a
large database with lung ultrasonographies with confirmed
diagnostic.
For now, there is a lot of room for improvements and tests.
The accuracy of the system must also be tested on the special
case of lung lesions, but for now we do not have such cases in
our database. From a theoretical point of view, this may result
a.
in another set of features that will need to be extracted from the
lesion contour which we consider for further developments.

77
 V. CONCLUSIONS REFERENCES
The work presented in this paper is the startup of a national [1] J. Didkowska, U. Wojciechowska, M. Manczuk, J. Lobaszewski, “Lung
research project which promises to improve the life expectancy cancer epidemiology: contemporary and future challenges worldwide”,
in Ann Transl Med. 2016;4(8):150. DOI: 10.21037/atm.2016.03.11
of lung cancer patients by early diagnosis. The main cause that
[2] DR. Brenner et all., “Previous lung diseases and lung cancer risk: a
prevents early diagnosis of lung cancer is that it can be pooled analysis from the International Lung Cancer Consortium.”, in
diagnosed only by CT of MRI which cannot be performed Am J Epidemiol. 2012 Oct 1;176(7):573-85. Epub 2012 Sep 17.
every year (as a check-up) without a doctor’s recommendation, [3] R.I. Chira et. all, „Transthoracic Ultrasonography: Advantages and
being invasive tests and also expensive. An alternative is Limitations in the Assessment of Lung Cancer”, InTech, 2017
represented by transthoracic ultrasounds which is a non- [4] A. Sebei et. all, “Hybrid PET/MRI co-segmentation based on joint fuzzy
invasive test and reduced costs. The only drawback is that the connectedness and graph cut”, in Comput Methods Programs Biomed.
results depend on the experience of the physician. In this paper, 2017 Oct;149:29-41. doi: 10.1016/j.cmpb.2017.07.006.
we propose a novel approach for automatic detection lung [5] T. Nelson et. all, Three-dimensional ultrasound, Lippincott Williams &
tumors based on ultrasound images, to assist young doctors or Wilkins, Philadelphia, 1999
ultrasound technicians in recognizing lung tumor and [6] LF. Herbert, “Drawbacks and Limitations of Computed Tomography”,
in Tex Heart Inst J. 2004; 31(4): 345–348.
differentiate them from condensations. Experimental results
[7] V. Caselles, R. Kimmel, G. Sapiro, Geodesic active contours.
prove that the effectiveness of the developed system. International Journal of Computer Vision, Volume 22, Issue 1, pp. 61-
79, 1997.
ACKNOWLEDGMENT
[8] L.K.Wee, E. Supriyanto, W. M. H. B. W. Mahmud, “Ultrasound Image
This work was supported by a grant of the Romanian Processing: And Its Application Using Matlab”, Lap Lambert Academic
National Authority for Scientific Research and Innovation, Publishing GmbH KG, 2011, ISBN 3845417838, 112 pp
CNCS – UEFISCDI, project number PN-III-P2-2.1-PED-2016- [9] F. Boutaouche and N. Benamrane, “Diagnosis of breast leasions using
0425, contract number 178PED/2017. the local Chan-Vese model, hierachical fuzzy partitioning and fuzzy
decision tree induction”, in Iranian Journal of Fuzzy Systems 2017
[10] S. N. Acho and W. I. D. Rae, „ Dependence of shape-based descriptors
and mass segmentation areas on initial contour placement using the
chan-vese method on digital mammograms”, in Computational and
Mathematical Methods in Medicine, 2015, 1-16.
[11] P. Rahmati et. all, “Mammography segmentation with maximum
likelihood active contours”, in Med. Img. Analysis, 2012, 1167–1186.

78