Original Article

Three-dimensional lung nodule

segmentation and shape variance
analysis to detect lung cancer with
reduced false positives

Proc IMechE Part H:

J Engineering in Medicine
2016, Vol. 230(1) 5870
IMechE 2016
Reprints and permissions:
sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/0954411915619951
pih.sagepub.com

Senthilkumar Krishnamurthy1, Ganesh Narasimhan2 and

Umamaheswari Rengasamy3

Abstract
The three-dimensional analysis on lung computed tomography scan was carried out in this study to detect the malignant
lung nodules. An automatic three-dimensional segmentation algorithm proposed here efficiently segmented the tissue
clusters (nodules) inside the lung. However, an automatic morphological region-grow segmentation algorithm that was
implemented to segment the well-circumscribed nodules present inside the lung did not segment the juxta-pleural
nodule present on the inner surface of wall of the lung. A novel edge bridge and fill technique is proposed in this article
to segment the juxta-pleural and pleural-tail nodules accurately. The centroid shift of each candidate nodule was computed. The nodules with more centroid shift in the consecutive slices were eliminated since malignant nodules resultant
position did not usually deviate. The three-dimensional shape variation and edge sharp analyses were performed to
reduce the false positives and to classify the malignant nodules. The change in area and equivalent diameter was more
for malignant nodules in the consecutive slices and the malignant nodules showed a sharp edge. Segmentation was followed by three-dimensional centroid, shape and edge analysis which was carried out on a lung computed tomography
database of 20 patient with 25 malignant nodules. The algorithms proposed in this article precisely detected 22 malignant
nodules and failed to detect 3 with a sensitivity of 88%. Furthermore, this algorithm correctly eliminated 216 tissue clusters that were initially segmented as nodules; however, 41 non-malignant tissue clusters were detected as malignant
nodules. Therefore, the false positive of this algorithm was 2.05 per patient.

Keywords
Computed tomography, juxta-pleural nodule, lung cancer, morphology processing, three-dimensional segmentation,
shape feature extraction

Date received: 7 June 2015; accepted: 3 November 2015

Introduction
Lung cancer is the leading cause of cancer-related
deaths globally.1 The diagnosis of lung cancer at early
stage is critical and uncertain as the physicians direct
the patient to undergo biopsy only after analyzing the
multiple lung computed tomography (CT) scans taken
between a time interval of 6 and 18 months.2 Nowadays,
although advanced imaging techniques such as CT scanning that precisely capture the images of lung are available, finding the cancerous nodules is still a challenging
task for physicians.3 CT scan of lung produces continuous cross-sectional images, and to confirm the cancerous
nature of lung, it is essential to analyze every cross-section. The radiologist needs to put extra efforts to analyze each cross-sectional image of lung and hence there

is high probability of an error. The development of

computer-aided diagnostic (CAD) tools will help physicians to more accurately analyze the CT scan images by
reducing the image reading time.
The small mass of tissue inside the lung, called
nodule, is an indicator of lung cancer. However, all

Department of ICE, Anna University, Chennai, India

Department of ECE, Rajalakshmi Institute of Technology, Chennai, India
3
Department of EEE, Velammal Engineering College, Chennai, India
2

Corresponding author:
Senthilkumar Krishnamurthy, Department of Electronics and
Communication, Rajalakshmi Institute of Technology, Chennai, Tamilnadu
600124, India.
Email: tkseneee@gmail.com

Downloaded from pih.sagepub.com by guest on January 12, 2016

Krishnamurthy et al.

59

lung nodules are not cancerous. Hence, quantitatively

defining and classifying the lung nodules is an interesting and challenging research task. The effective segmentation techniques need to be applied on lung CT
scan images to segment the lung nodules.
The pixel intensitybased lung nodule segmentation
has been used to segment lung nodules in many studies.46 In these studies, based on the knowledge of CT
scanning technology and X-ray dose, the threshold has
been fixed for nodules in terms of Hounsfield unit
(HU). Morphological processing needs to be performed
in threshold-based segmentation algorithms to detect
the candidate nodule by eliminating background noise
of lung parenchyma, fat and bone parts. Model-based
template matching methods are also used for segmenting of the lung nodule.79 The disadvantage of template
matching technique is that it shows more false positives
(FPs). The two-dimensional (2D) analysis of lung
nodule segmentation algorithms shows more FP results
than the three-dimensional (3D) analysis. Hence, 3D
analysis is used to reduce FPs. The work implemented
by Alilou et al.10 mainly focused on 3D structural visualization of lung nodules which produced 3.9 FPs per
patient. The work published by Demir and Ylmaz
Cxamurcu11 dealing with the texture 3D model of lung
nodules resulted in a FP of 2.45 per patient. The work
published by Lu et al.12 using a complex hybrid model
for nodule segmentation and classification resulted in a
FP of 3.13 per patient. The genetic and fuzzy rulebased
algorithm proposed by Ozekes et al.13 results in 100%
sensitivity but with the cost of 13.4 FPs per patient.
In this study, we proposed the segmentation algorithms that automatically select the initial threshold
based on the anatomical structure of lung, and then
from the initial threshold seed point, it starts to grow
and segment the different regions of lung. The 3D analysis was performed by analyzing the consecutive sections for every candidate nodule that was segmented
with the 2D analysis. We modeled the 3D structural
behavior of nodule candidates quantitatively in terms
of centroid shift, change in area, change in equivalent
diameter and edge sharp index (ESI), which reduced
the FPs to great extent. This step makes this work different and efficient from other works in the literature
which were discussed above.
Nodules are the clusters of tissue in the parenchyma
of lung. Based on its position in the parenchymal region,
they are classified under four categories14 as shown in
Figure 1: (A) well-circumscribed nodules are discrete,
well-marginated, with rounded opacity, and completely
surrounded by lung parenchyma; (B) pleural-tail nodules
are situated near the pleural surface and connected by
thin linelike structure; (C) vascularized nodules are significantly connected with neighboring vessels and located
centrally in the lung; and (D) juxta-pleural nodules are
significantly connected with pleural surface. In this work,
we proposed the segmentation algorithms that can segment nodules located within the parenchyma as well at
the pleural surface.

Figure 1. Template model of lung nodule.

Segmenting the lung nodules from the CT scan is

really a challenging task because of the following
reason:
1.
2.
3.
4.

5.

The pixel intensities of different regions in the lung

CT scan overlaps.
Manual threshold setting should be avoided to
make the segmentation process automatic.
The shape of the nodules present inside parenchyma region is not uniform.
The position of the nodules inside the lung makes
the segmentation process tedious. Developing a
common algorithm to segment well-circumscribed,
pleural tail, vascularized and juxta-pleural nodules
makes the segmentation work challenging.
Segmentation needs to be carried out continuously on consecutive slices to monitor the behavior of nodule candidates from one slice to the
other.

Methods and materials

SPIE American Association of Physicists in Medicine
(AAPM) lung challenge database15 and Lung Image
Data Consortium (LIDC) database16 were used in this
work. The SPIE AAPM database has lung CT scans
of 10 patients with the ground truth information about
nodules position in separate excel sheet. Whereas the
LIDC database has lung CT slices of 1001 patients
along with four radiologists ground truth report. In
this study, 10 cases of SPIE and 10 cases of LIDC were
used. The number of slices in each CT scan of this
database was varied from 110 to 388. A total of 4896
sections of CT images from 20 cases were considered
for this work. Each image was 512 3 512 in size with
12-bit gray resolution for SPIE AAPM and 16-bit gray
resolution for LIDC database. For the 20 cases
involved in this study, the scan images were taken with
the help of Philips or GE Medical systems. In this
database of 20 patient, 25 malignant nodules are present, in which 14 nodules are well circumscribed, six are
juxta pleural, three are pleural tail and two are
vascularized.

Downloaded from pih.sagepub.com by guest on January 12, 2016

60

Proc IMechE Part H: J Engineering in Medicine 230(1)

Figure 2. Lung segmentation masks: (a) before applying region

grow, (b) region grow mask, (c) inverted region grow mask and
(d) after opening of small clusters.

Figure 3. Segmented lung lobe and nodule image: (a) lung lobe
mask, (b) original lung lobe, (c) nodules and vessels mask and (d)
candidate nodules.

Automatic morphological region-grow segmentation

for well-circumscribed nodule
The threshold-based segmentation methods are not
fully automatic as the thresholds need to be fixed after
analyzing the given lung CT image.1720 We have developed a complete automatic algorithm to segment the
nodule structures from the CT scan sections. The
region growing was applied in such a way that the initial seed position can be automatically decided by the
algorithm. Morphological operations were applied on
the region grown to mask efficiently and segment the
well-circumscribed nodule from the lung parenchyma
in the CT images.
The algorithm steps used were as follows:
1.

2.

The initial seed position was automatically chosen

as [(m/2), (n/2)], where m is the number of rows
and n is the number of columns of the CT image.
From the anatomical structure of lungs, in the CT
scan image, the heart appears in between the two
lungs (Figure 2(a)), which is the central region of
lung CT scan slice. The algorithm was developed
to check the pixel intensity value at (m/2, n/2)
before it starts growing. If the seed position value
is close to black region due to vessel hole, the initial seed was shifted for few positions in x and y
before region grow was started
8 h m ni
<
, ,
if I(S) is high
2 2
i
1
S = hm
n
:
+ rx, + ry , if I(S) is low
2
2
Region growing was performed with an automatically chosen seed in the step 1, which created the
mask for the muscle/fat portion of lung CT slice as
1 and remaining portion as 0 (Figure 2(b)).

3.

4.

5.

6.

I(S) chosen in step one compare with all its eight

neighbor. The neighbor pixels with less difference
made as 1 and other pixels changed to 0. This process iteratively grows for every new 1 pixel position produced in last iteration until no more pixel
difference is minimum.
The image mask produced in step 2 was inverted
as shown in Figure 2(c), since we need to locate the
region of lung lobe without the muscle and fat
portion.
In the inverted image mask produced in step 3 the
lung lobes appeared as white region along with
some major central vessels and background region
as white too. To remove small white regions which
appeared due to vessel cavities, the morphological
opening was performed. All the white regions with
less connected white pixels were made as black as
shown in Figure 2(d).
The remaining white regions in the mask after performing step 4 were lung lobes along with background and any other unwanted regions. All the
connected white regions were counted and labeled.
The area and eccentricity of the each connected
white region were computed, and the region with
maximum area and/or eccentricity deviating from
the value 1 was removed. The background area
was more compared with the lung lobes and
unwanted white cluster regions along with background that has deviated eccentricity value from 1.
Thus, after step 5 only the lung lobes were segmented from the entire lung CT scan slice as shown in
Figure 3(a).
The lung lobe was masked in the original lung CT
scan slice, so that the parenchymal region alone
was segmented successfully (Figure 3(b)). The wellcircumscribed nodules like structures appeared
inside the parenchyma region. The lung parenchyma has shown dark background with bright
nodules and vessels that were spread inside it.

Downloaded from pih.sagepub.com by guest on January 12, 2016

Krishnamurthy et al.

61

Therefore, by simply converting this into binary,

the suspected nodules and vessels part became 1,
and all background dark portions became 0. Thus,
the mask of suspected lung nodules along with vessel lines was segmented as in Figure 3(c). Based on
eccentricity and area criteria, very small tissue cluster and the vessels were removed as shown in
Figure 3(d).

Juxta-pleural nodule segmentation using edge bridge

and fill technique
Juxta-pleural nodules are significantly connected with
the pleural surface of the lung, that is, the inside wall
portion of the lung as shown in Figure 4(a). As juxta
nodules were connected with the wall portion, while we
removed the wall portion by any connected componentbased algorithms, these juxta-pleural nodules were also
removed from the parenchyma region. Hence, it is essential to do modifications in the automatic morphological

region-grow segmentation (AMRG) algorithm as discussed in the previous section. Applying AMRG on the
juxta-pleural nodule CT slice will produce the initial
mask as shown in Figure 4(b).
In Figure 4(b), the juxta-pleural nodule at the
parenchyma-wall surface was removed, but it was
essential to keep that nodule portion. Usually morphological filling can be performed to fill the holes (black
region) in the place of juxta-pleural nodule, but since
this hole was not surrounded by the white pixels, the
filling operation may have filled all background black
regions as white. Therefore, it was essential to connect
the broken-edge portion of the lung wall around the
juxta-pleural hole.
Novel edge bridge and fill technique (EBFT) was
used to bridge the gaps in the edges of inner wall of
lung. The steps of EBFT were as follows:
1.

2.

3.

Figure 4. Juxta nodule: (a) CT slice with juxta nodule and (b)
lobe mask without juxta nodule.

The mask was created using the steps 1, 2 and 3 of

AMRG algorithms as discussed in previous
section.
Only extreme boundary line of the mask was kept
and all the inner portions were removed as shown
in Figure 5(a). The juxta hole (black region) is connected to the background black region; hence, it is
essential to separate the juxta nodule hole region
from the background. The juxta nodule hole edge
in Figure 5(a) needs to be closed, so that the nodule
region can discriminate from the background.
To close this open-edge portion of juxta nodule, an
iterative thickening (dilation) and thinning (skeletenization) morphology21,22 is used in this work.
Dilation is performed with a disk or square structural element of size 4. The dilation of image A by
the structuring element B is defined by equation (2)

Figure 5. EBFT segmentation: (a) lung boundary, (b) dilated closed lung boundary, (c) skeletonized closed boundary, (d) closed lung
mask, (e) hole-filled lung mask, (f) lung region with juxta-pleural nodule, (g) binary candidate nodules and (h) gray candidate nodules.

Downloaded from pih.sagepub.com by guest on January 12, 2016

62

Proc IMechE Part H: J Engineering in Medicine 230(1)

2

A xor B = [ Ab
b2B

Disk structural element is preferable over square as

edge growing is smooth with disk element. Initially, in
this work, the structural element of size 12 is used and
all juxta nodules in this database are segmented as
shown in Figure 5, but still to make the juxta nodule
segmentation more robust and reliable for any size of
nodule, it is preferable to use small size structural element iteratively rather than using big structural element
once. In this work, the disk structural element with
radius 4, as shown in Figure 6, is used followed by skeletenization iteratively for three times. This process is
independent of the number of iteration occurring after
the closing of juxta nodule edge, hence the maximum
limit of iterations does not affect the reliability of this
juxta nodule segmentation process. The juxta nodule
hole, edge pattern in iterative steps and segmented
nodule are shown in Figure 7.
4.

The dilated thick line from equation (2) was skeletonized by removing the pixels from each side with

5.

6.

respect to the central pixels of the line until there

were no pixels to remove on the either side of the
boundary.23 This made the boundary lines more
sharp and without gap as shown in Figure 5(c).
The iterative dilation followed by skeletenization is
shown in Figure 7(c)(h).
This closed boundary line was added to the mask
created in step 1 to close the parenchyma-wall
boundary without any gap as shown in Figures
5(d) and 7(i).
The holes inside the parenchyma including juxta
pleural were filled as shown in Figure 5(e). Thus,
the holes filled mask with the original CT slice
gave the parenchyma region with the juxta-pleural
nodules as shown in Figure 5(f). Then the image in
Figure 5(f) was converted to binary to mask the
nodule as shown in Figure 5(g). Furthermore, the
nodule mask with the original image gave the
original texture version of the nodules as in
Figure 5(h).

EBFT technique can also segment pleural-tail

nodules, well-circumscribed and vascularized nodules.

3D centroid-shift analysis to classify nodules from

vessels

Figure 6. Disk structural element pattern for dilation.

The tissue clusters segmented using AMRG and EBFT

algorithms contain both nodules and non-nodular
structures. Some calcified patterns and vessels from 2D
slices of CT scan segmented by the AMRG and EBFT
algorithms are actually not nodules, but these are the
tissue clusters which appear like nodules. It is very

Figure 7. Juxta nodule edge bridging: (a) juxta nodule hole, (b) juxta nodule edge, (c) first iteration dilated edge, (d) first iteration
skeletonized edge, (e) second iteration dilated edge, (f) second iteration skeletonized edge, (g) third iteration dilated edge, (h) third
iteration skeletonized edge, (i) closed juxta nodule edge, (j) filled juxta nodule hole, (k) binary juxta nodule and (l) gray juxta nodule.

Downloaded from pih.sagepub.com by guest on January 12, 2016

Krishnamurthy et al.

63

essential to find the feature to discriminate non-nodules

from candidate nodules.
The vessels which are segmented incorrectly as
nodules can be eliminated by analyzing the consecutive
CT slices. From one slice to another, the vessel clusters
position changes drastically, whereas the actual nodule
position remains relatively same. Hence, by comparing
the centroid values of the segmented clusters from one
slice to other, vessels can be eliminated from candidate
nodules.24,25
The algorithmic steps used for centroid-shift analysis
were as follows:
1.

The centroid value of each segmented cluster (suspected nodules) from the first CT slice was determined using the equation
Cx, Cy =
" PM PN

i=1
j = 1 i  fi, j
,
PM PN
i=1
j = 1 fi, j

6.

The non-cancerous benign nodule will not grow in

size. The growth of the nodule can be analyzed by
comparing the scans of the same patient taken at
different time intervals (6 months to 2 years), based
on patients smoking habits and environmental
conditions.

We evaluated the uniformity in the shape of segmented nodules in the consecutive slices and classified them
as either benign or malignant. Also, the ESI was computed; the spiked edge is an indication of the malignant
nodule.
The area, eccentricity, equivalent diameter and ESI
were calculated for each nodule candidates from each
slice using the following equations28,29
Area =

PM

PN
i=1
j = 1 j  fi, j
PM PN
i=1
j = 1 fi, j

m X
n
X

fbx, y

x=1 y=1

 Pm

max
x = 1 fbx, :
nP
o
Eccentricity =
n
max
y = 1 fb( : , y)

where f(i, j) is a binary image with M rows and N

columns.

ED =

2 p
area
=P

2.

ESI =

magnitudegradientfi, j


length xgradient

3.

The centroid values were computed for next three

to four consecutive CT slice segmented clusters.
The clusters were analyzed further if the centroid
shift was found to be minimum.

3D shape and edge analysis to differentiate

malignant nodules from benign nodules
Although the vessels that appeared like nodule in 2D
analysis were removed, there were also remaining segmented objects other than the cancerous lung nodules.
These non-cancerous lung nodules are known as nonnodules, calcifications or benign nodules. Benign are
formed in lung due to inflammation and scarring of
lung tissue associated with bacterial or fungal infection.
A type of nodule that is unlikely to be cancer is known
as calcified granuloma which consists of calcium
deposits.
The characteristics of benign lung nodules26,27 are as
follows:
1.
2.
3.

4.

5.

They are smaller in size ( \ 5 mm).

The nodules between 5 and 10 mm need further
investigation based on its shape and texture.
Benign nodules may contain calcium deposits,
which are visible as uniform bright spot in CT
scan.
The probability of lung nodule being benign
increases with size and regularity in its shape. The
more regular the shape of the nodule, the more
likely it is benign.
Almost all the nodules that are not cancerous possess very smooth or rounded edges.

If for a nodule object, from one slice to other, area

and equiv diameter values remained unchanged, then
the chances of nodule being benign or calcification are
more.

Results
The candidate nodules, after applying the AMRG and
EBFT algorithm for two different cases, are shown in
Figures 3(d) and 5(g). In these images, there are many
candidate nodules of . 3-mm size.
The radiologist report for these images indicated
that only one malignant nodule exists in each case. To
remove the FPs, 3D position and shape analysis have
been carried out. The centroid values of each candidate
nodule were compared with next consecutive three to
four slices. If the shift in the centroid of a particular
candidate nodule in consecutive slices is less, then the
chance for that structure to be nodule is more. All candidate nodules with high variation in the centroid
between consecutive slices were eliminated. Figure 8
shows the nodules from consecutive CT slices whose
centroids did not vary a lot. Based on its centroid shift,
many suspected nodules were eliminated from Figure
3(d) and only two nodules showed minimum centroid
shift in consecutive CT slices. The nodules shown in
Figure 8 were present in four consecutive slices. The
area, equiv diameter and ESI were computed for a
nodule in each slice until it vanishes from the CT scan
slice, and the values are tabulated in Table 1 for two
different candidate nodule series.

Downloaded from pih.sagepub.com by guest on January 12, 2016

64

Proc IMechE Part H: J Engineering in Medicine 230(1)

Figure 8. Nodules from consecutive slices with minimum centroid shift (case 1).

Table 1. Shape variation features of case 1.

Nodule candidate 1

Nod1_Slice1
Nod1_Slice2
Nod1_Slice3
Nod1_Slice4

For four consecutive

slices in
which same nodule
structure exists

Nodule candidate 2
Area

Equiv
diameter

Edge sharp
index

40
75
116
131

7.14
9.77
12.15
12.9

0.136
0.151
0.241
0.211

Change
in area

Change in
equiv diameter

35
41
15

2.63
2.33
0.8

Nod2_Slice1
Nod2_Slice2
Nod2_Slice3
Nod2_Slice4

For four consecutive slices

in which same nodule
structure exists

The data in Table 1 are plotted in Figure 9(a)(c),

which graphically indicated how the nodules area,
equiv diameter and edge sharpness vary from one slice
to other. The values of area, equiv diameter and ESI
are tabulated in Tables 24 for the nodule candidates,
which segmented using EBFT algorithm for a juxtapleural case CT slices. In this case, after applying the
3D centroid-shift analysis, five nodules showed minimum centroid shift in consecutive slices as shown in
Figure 10.
Generally, the calcium deposits accumulated inside
lung parenchyma are wrongly segmented as nodule candidates. The calcification cluster from one case of SPIE
database is shown in Figure 12, and the shape parameters are tabulated in Table 5.

Discussion
In Figure 8, although the FPs were reduced to great
extent, one of two candidate nodules were reported as
malignant by the radiologist. Therefore, additional 3D

Area

Equiv
diameter

Edge sharp
index

34
32
36
33

6.58
6.57
6.58
6.58

0.1
0.08
0.097
0.108

Change
in area

Change in equiv
diameter

2
4
3

0.01
0.01
0

morphological analysis was carried out. The different

morphological features such as area, perimeter, eccentricity ratio, equiv diameter, solidity, extent, and edge
sharp index were computed. The nodule for which the
area and equiv diameter values were not much changed
from one slice to other was marked as non-malignant
by the radiologist. In addition, the edge sharp index
value was more for a candidate nodule candidate that
was marked as malignant by the radiologist.
For case 1 from the database, series of scan slices
were taken and segmentation algorithm was applied on
each scan. The FPs were reduced by centroid-shift analysis and finally two candidate nodules remained. The
candidate nodule 1 was present from second to fifth
slice, whereas candidate nodule 2 was present from
fourth to seventh slice; both these nodules were present
in four consecutive slices. The change in the nodule area
and equiv diameter between the slices were computed
and tabulated. For candidate nodule 1, the change in
area and equiv diameter between consecutive slices was
more than the candidate nodule 2. This suggested that

Downloaded from pih.sagepub.com by guest on January 12, 2016

Krishnamurthy et al.

65

Figure 9. Shape variation plots: (a) nodule area variation, (b) nodule equiv diameter variation and (c) nodules edge sharp plot.

Table 2. Shape variation features of case 2 (nodules 1 and 2).

Nodule candidate 1

Nod1_Slice1
Nod1_Slice2
Nod1_Slice3
Nod1_Slice4
Nod1_Slice5
Nod1_Slice6

For six consecutive slices

in which same nodule
structure exists

Nodule candidate 2
Area

Equiv
diameter

Edge sharp
index

749
879
912
902
880
802

30.8
33.5
34.1
33.8
33.4
31.9

0.553
0.649
0.683
0.63
0.62
0.67

Change
in area

Change in equiv diameter

130
33
10
22
78

2.7
0.6
0.3
0.4
1.5

Area

Equiv
diameter

Edge sharp
index

Nod2_Slice1

31

6.28

0.087

Nod2_Slice2

26

5.75

0.0766

Nod2_Slice3

28

5.9

0.0874

Change
in area

Change in equiv diameter

0.53

0.15

For three consecutive

slices in which
same nodule
structure exists

Downloaded from pih.sagepub.com by guest on January 12, 2016

66

Proc IMechE Part H: J Engineering in Medicine 230(1)

Table 3. Shape variation features of case 2 (nodules 2 and 3).

Nodule candidate 3

Nod3_Slice1
Nod3_Slice2
Nod3_Slice3
Nod3_Slice4
Nod3_Slice5
Nod3_Slice6

For six consecutive slices

in which same nodule
structure exist

Nodule candidate 4
Area

Equiv
diameter

Edge sharp
index

25
23
23
26
22
19

5.64
5.41
5.4
5.75
5.3
4.9

0.082
0.073
0.078
0.07
0.074
0.059

Change
in area

Change in equiv diameter

2
0
3
4
3

0.23
0.01
0.35
0.45
0.4

Nod4_Slice1
Nod4_Slice2
Nod4_Slice3
Nod4_Slice4
Nod4_Slice5

For five consecutive

slices in which same
nodule structure exist

Area

Equiv
diameter

Edge sharp
index

40
37
40
35
33

7.13
6.8
7.13
6.67
6.4

0.109
0.106
0.1
0.104
0.084

Change
in area

Change in equiv diameter

3
3
5
2

0.33
0.33
0.46
0.27

Table 4. Shape variation features of case 2 (nodule 5).

Nod5_Slice 1
Nod5_Slice2
Nod5_Slice3

For three consecutive slices in which same nodule structure exists

Area

Equiv diameter

Edge sharp index

43
34
27

7.39
6.57
5.8

0.12
0.08
0.076

Change in area

Change in equiv diameter

9
7

0.82
0.77

Figure 10. Nodules from consecutive slices with minimum centroid shift (case 2).

Downloaded from pih.sagepub.com by guest on January 12, 2016

Krishnamurthy et al.

67

Figure 11. Shape variation plots for case 2: (a) nodule area variation, (b) nodule equiv diameter variation and (c) nodules edge
sharp plot.

Figure 12. Calcification patterns from consecutive slices.

the shape of the nodule 1 was varying from one slice to

another, whereas the shape of the nodule 2 was uniform
across the slices. Therefore, the nodule 1 was more
likely to be malignant and nodule 2 was more likely to
be benign. Additionally, the ESI value was also more
for nodule 1 slices in Table 1, which indicates that the
nodule 1 has sharp edge. The probability of malignancy
is more when the nodule has sharp edge. In Table 1, the
lower value of ESI for candidate nodule 2 indicates that
the edges of the nodule are smooth. Therefore, the possibility of nodule 2 being benign or calcification is more.

Table 5. Shape feature of calcification.

Nodule candidates from

six consecutive slices

ESI: edge sharp index.

Downloaded from pih.sagepub.com by guest on January 12, 2016

Area

Equiv diameter

ESI

54
52
50
49
48
50

8.3
8.1
7.9
7.7
7.8
7.9

0.099
0.097
0.100
0.098
0.101
0.097

68

Proc IMechE Part H: J Engineering in Medicine 230(1)

The radiologist report for this patient confirmed that

the candidate nodule 1 is malignant. This matches with
the K-means Euclidean distancebased clustering carried out on data from Table 1.
Similarly, the change in area, change in equiv diameter and ESI from consecutive slices for case 2 are
tabulated in Tables 35 and plotted in Figure 11. In this
case, there are five candidate nodules of which centroid
shift is minimum. The change in area for nodule 1 from
Table 3 is plotted as red contour in Figure 11(a). The
change in area for nodule 1 is much higher compared
with all other four nodules as depicted in Figure 11(a).
Furthermore, the variation in equiv diameter is also
high compared with all other nodules as plotted in
Figure 11(b). For nodule 1, the ESI value is more than
all other nodules, which indicates that nodule 1 has
sharp edge. The area and equiv diameter varied much
in consecutive slices for nodule 1, indicating that the
nodule 1 has irregularity in shape between consecutive
slices compared with all other nodules. The K-means
Euclidean distancebased clustering on data from
Tables 24 distinguished the nodule 1 as separate cluster
and combined remaining four nodules as another cluster. Therefore, as per the data tabulated, plotted and
clustered, the possibility of nodule 1 being malignant is
more. This was further confirmed by radiologist.
Table 6. Statistical results.
Statistic
Sensitivity
Specificity
Positive likelihood ratio
Negative likelihood ratio
Disease prevalence

Formula
TP
TP + FN
TN
TN + FP
Sensitivity
100  Specificity
100  Sensitivity
Specificity
TP + FN
TP + FN + TN + FP

Value
88%
84.05%
5.51
0.142
8.87%

The calcium deposits accumulated (calcifications)

inside lung are not cancerous in nature. Hence, these
calcifications should be eliminated; otherwise, the FPs
will increase. Texture and edge-based analysis30 will
discriminate these calcium deposits from malignant
nodules as the contrast of calcifications is more. In this
work, shape-based analysis is carried out to identify the
calcifications. The calcium deposits are uniform in
shape in the consecutive slices of lung CT scan, and its
edge (boundary) is very smooth. The calcification patterns of one patient case from database are shown in
Figure 12. Table 5 shows that an area and equiv diameter of this calcification are not varying much in consecutive slices and also the ESI value is less, which
means the nodule candidate shape is uniform and an
edge is smooth. Therefore, this nodule candidate shown
in Figure 12 is calcification. Similarly, all the calcifications were detected and eliminated, which greatly
reduce the FPs of this work.
The segmentation algorithm (AMRG and EBFT)
developed in this article was applied to 10 cases from
SPIE and 10 cases from LIDC database. For these 20
cases, the candidate nodules with a size of . 3 mm were
segmented successfully. The 3D centroid, shape and
edge analysis have been carried out on the nodule candidates. Out of 25 malignant nodules present in this
database, the algorithms implemented in this article
detected 22 malignant nodules and failed to detect 3.
Therefore, the true positive (TP) score is 22 and false
negative (FN) score is 3. This algorithm also detected
41 nodules from segmented nodule candidate as malignant which is not marked as malignant nodule by the
radiologists and correctly eliminated 216 nodules candidates; therefore, the FP is 41 and the true negative (TN)
is 216. Table 6 gives the performance measure of this
work. The FP per case for this work is 2.05 with a good
sensitivity of 88%. A comparative account of the performance of our method with other methods3136 is
given in Table 7.

Table 7. Performance comparison of reported CAD system.

CAD system

Sensitivity (%)

False positive per patient

Nodule size criteria (mm)

Zhao5
Opfer and Wiemeker31
Dehmeshki et al.33
Ozekes et al.13
Golosio et al.36
Suarez-Cuenca et al.34
Choi WJ and Choi TS35
Alilou et al.10
Demir and Ylmaz C
xamurcu11
12
Lu et al.
Our framework

84
74
90
100
79
80
95
80
98
85
88

5
4
15.5
13.4
4
7.7
2.27
3.9
2.47
3.13
2.05

27
.4
320
.5
.4
.4
. 1.5
.4
NA
.5
.3

CAD: computer-aided diagnosis.

Downloaded from pih.sagepub.com by guest on January 12, 2016

Krishnamurthy et al.

69

Conclusion
An efficient and completely automatic segmentation
followed by the 3D morphology analysis to classify the
malignant lung nodules from CT scan was successfully
implemented in this study. The AMRG algorithm
developed, segmented the well-circumscribed nodules,
but was unable to segment the juxta-pleural nodules.
This drawback was overcome by the development of
novel EBFT algorithm implemented in this work. The
3D centroid analysis followed by 3D shape variance
and ESI analysis, carried out on consecutive slices that
remarkably reduced the FPs. This algorithm was
applied on 20 cases having a total of 25 malignant
nodules. A total of 22 of 25 nodules were correctly
detected and 4 nodules were not detected. Of the 257
segmented candidate nodules (excluding 21 malignant
nodules), 216 nodules were eliminated correctly and 41
candidate nodules were incorrectly detected as malignant nodules. This work produced the better result in
terms of 2.05 FP per case with sensitivity of 88%. This
can be improved further in future work by analyzing
the texture of candidate nodules along with the morphological variations.
Acknowledgements
We thank The Cancer Imaging Archive (TCIA) for
making the LIDC and SPIE AAPM database available
in online for open access. Proper references are quoted
for the database used in this work from TCIA.

6.

7.

8.

9.

10.

11.

12.

13.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest
with respect to the research, authorship and/or publication of this article.

14.

Funding
The author(s) received no financial support for the
research, authorship and/or publication of this article.

15.

16.

References
1. American Cancer Society. Cancer facts & figures 2014.
Atlanta, GA: American Cancer Society, 2014.
2. Gould MK, Donington J, Lynch WR, et al. Evaluation
of individuals with pulmonary nodules: when is it lung
cancer? Diagnosis and management of lung cancer:
American college of chest physicians evidence-based clinical practice guidelines. Chest 2013; 143(Suppl. 5): e93S
e120S.
3. Chheang S and Brown K. Lung cancer staging: clinical
and radiologic perspectives. Semin Intervent Radiol 2013;
30(2): 99113.
4. Armato SG III, Giger ML and MacMahon H. Automated detection of lung nodules in CT scans: preliminary
results. Med Phys 2001; 28(8): 15521561.
5. Zhao B, Gamsu G, Ginsberg MS, et al. Automatic detection of small lung nodules on CT utilizing a local density

17.

18.

19.

maximum algorithm. J Appl Clin Med Phys 2003; 4(3):

248260.
Shao H, Cao L and Liu Y. A detection approach for
solitary pulmonary nodules based on CT images. In: Proceedings of the 2012 2nd international conference on computer science and network technology (ICCSNT),
Changchun, China, 2931 December 2012, pp.1253
1257. New York: IEEE.
Elhabian SY, El Munim HA, Elshazly S, et al. Experiments on sensitivity of template matching for lung
nodule detection in low dose CT scans. In: Proceedings
of the 2007 IEEE international symposium on signal processing and information technology, Giza, 1518 December 2007, pp.10291035. New York: IEEE.
Nagata R, Kawaguchi T and Miyake H. Automated
detection of lung nodules in chest radiographs using a
false-positive reduction scheme based on template matching. In: Proceedings of the 2012 5th international conference on biomedical engineering and informatics (BMEI),
Chongqing, China, 1618 October 2012, pp.216223.
New York: IEEE.
Jo HH, Hong H and Goo JM. Pulmonary nodule registration in serial CT scans using global rib matching and
nodule template matching. Comput Biol Med 2014; 45:
8797.
Alilou M, Kovalev V, Snezhko E, et al. A comprehensive
framework for automatic detection of pulmonary nodules
in lung CT images. Image Anal Stereol 2014; 33(1): 1327.
Demir O and Ylmaz C
xamurcu A. Computer-aided detection of lung nodules using outer surface features. Biomed
Mater Eng 2015; 26: 12131222.
Lu L, Tan Y, Schwartz LH, et al. Hybrid detection of
lung nodules on CT scan images. Med Phys 2015; 42:
50425054.
Ozekes S, Osman O and Ucan ON. Nodule detection in
a lung region thats segmented with using genetic cellular
neural networks and 3D template matching with fuzzy
rule based thresholding. Korean J Radiol 2008; 9(1): 19.
Kostis WJ, Reeves AP, Yankelevitz DF, et al. Threedimensional segmentation and growth-rate estimation of
small pulmonary nodules in helical CT images. IEEE
Trans Med Imaging 2003; 22(10): 12591274.
Armato SG III, Hadjiiski L, Tourassi GD, et al. Guest
editorial: LUNGx challenge for computerized lung
nodule classification: reflections and lessons learned. J
Med Imag 2015; 2(2): 020103.
Armato SG III, McLennan G, Bidaut L, et al. The Lung
Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference
database of lung nodules on CT scans. Med Phys 2011;
38: 915931.
Senthil Kumar TK, Ganesh EN and Rengasamy U.
Automatic lung nodule segmentation using autoseed
region growing with morphological masking (ARGMM)
and feature extraction through complete local binary Pattern and microscopic information pattern. EuroMediterr
Biomed J 2015; 10(5): 99119.
Lassen BC, Jacobs C, Kuhnigk JM, et al. Robust semiautomatic segmentation of pulmonary subsolid nodules
in chest computed tomography scans. Phys Med Biol
2015; 60(3): 13071323.
Kumar TS and Ganesh EN. Proposed technique for accurate detection/segmentation of lung nodules using spline
wavelet techniques. Int J Biomed Sci 2013; 9(1): 917.

Downloaded from pih.sagepub.com by guest on January 12, 2016

70

Proc IMechE Part H: J Engineering in Medicine 230(1)

20. Kumar TS, Narasimhan G and Rengasamy U. Texture

Pattern Based Lung Nodule Detection (TPLND) technique in CT images. IRECOS 2014; 9(3): 415426.
21. Bhateja V and Devi S. A novel framework for edge detection of microcalcifications using a non-linear enhancement operator and morphological filter. In: Proceedings
of the 2011 3rd international conference on electronics computer technology (ICECT), vol. 5, Kanyakumari, India,
810 April 2011. New York: IEEE.
22. Dougherty ER and Lotufo RA. Hands-on morphological
image processing, vol. 71. Bellingham, WA: SPIE Press
(International Society for Optical Engineering), 2003.
23. Saeed K, Rybnik M and Tabedzki M. Implementation
and advanced results on the non-interrupted skeletonization
algorithm (Lecture Notes on Computer Science). Berlin,
Heidelberg: Springer, 2001, pp.601609.
24. Li X, Wang X, Dai Y, et al. Supervised recursive segmentation of volumetric CT images for 3D reconstruction of
lung and vessel tree. Comput Methods Programs Biomed
2015; 122: 316329.
25. Weiss E, Wijesooriya K, Dill SV, et al. Tumor and normal tissue motion in the thorax during respiration: analysis of volumetric and positional variations using 4D CT.
Int J Radiat Oncol Biol Phys 2007; 67(1): 296307.
26. Ost DE and Gould MK. Decision making in patients
with pulmonary nodules. Am J Respir Crit Care Med
2012; 185(4): 363372.
27. MacMahon H, Austin JH, Gamsu G, et al. Guidelines
for management of small pulmonary nodules detected on
CT scans: a statement from the Fleischner Society. Radiology 2005; 237(2): 395400.
28. Jurie F and Schmid C. Scale-invariant shape features for
recognition of object categories. In: Proceedings of the
2004 IEEE computer society conference on computer vision
and pattern recognition, 2004 (CVPR 2004), vol. 2,

29.

30.

31.

32.

33.

34.

35.

36.

Washington, DC, 27 June2 July 2004, pp.II-90II-96.

New York: IEEE.
Yang M, Kpalma K and Ronsin J. A survey of shape feature extraction techniques. Pattern Recogn 2008: 4390.
https://hal.archives-ouvertes.fr/hal-00446037/document
Diaz-Huerta CC, Felipe-Riveron EM and MontanoZetina LM. Quantitative analysis of morphological techniques for automatic classification of micro-calcifications
in digitized mammograms. Expert Syst Appl 2014; 41(16):
73617369.
Opfer R and Wiemker R. Performance analysis for
computer-aided lung nodule detection on LIDC data. In:
Jiang Y and Sahiner B (eds) Medical imaging. Bellingham, WA: SPIE (International Society for Optics and
Photonics), 2007, pp.65151C.
Dolejs M and Kybic J. Automatic two-step detection of
pulmonary nodules. In: Jiang Y and Sahiner B (eds) Medical imaging. Bellingham, WA: SPIE (International Society for Optics and Photonics), 2007, pp.65143J.
Dehmeshki J, Ye X, Lin X, et al. Automated detection of
lung nodules in CT images using shape-based genetic
algorithm. Comput Med Imaging Graph 2007; 31(6): 408
417.
Suarez-Cuenca JJ, Tahoces PG, Souto M, et al. Application of the iris filter for automatic detection of pulmonary
nodules on computed tomography images. Comput Biol
Med 2009; 39(10): 921933.
Choi WJ and Choi TS. Automated pulmonary nodule
detection system in computed tomography images: a hierarchical block classification approach. Entropy 2013;
15(2): 507523.
Golosio B, Masala GL, Piccioli A, et al. A novel multithreshold method for nodule detection in lung CT. Med
Phys 2009; 36(8): 36073618.

Downloaded from pih.sagepub.com by guest on January 12, 2016