Australasian Physical & Engineering Sciences in Medicine

https://doi.org/10.1007/s13246-018-0678-z

SPECIAL ISSUE ARTICLE

Automatic segmentation of cervical region in colposcopic images

using K-means
Bing Bai1 · Pei‑Zhong Liu1,3 · Yong‑Zhao Du1,3 · Yan‑Ming Luo2

Received: 29 May 2018 / Accepted: 20 August 2018

© Australasian College of Physical Scientists and Engineers in Medicine 2018

Abstract
Colposcopy is an important imaging modality for the detection of cervical lesions. The analysis of colposcopic images,
especially the effective segmentation of the cervical region, has important clinical value in clinical application. A cervical
segmentation method based on the HSV color mode is proposed, which can divide and extract the cervical region in the
medical and anatomical sense. Firstly, the histogram threshold method is used to analyze the histogram (Y) of the colposcopic
image. In order to achieve the removal of the mirror reflection pretreatment operation in the colposcopy image. Secondly, the
Preprocessed RGB images is used. Then, the colposcopic image is converted into the HSV color space, and the V component
is extracted using the K-means algorithm. Finally, using the area filter to smooth the edge, the segmented cervical region
can be obtained. In our study, 110 standard colposcopy images, which were tagged by experts, were tested and verified.
The segmentation results were analyzed and compared using dice coefficients, Jaccard coefficients, structural segmentation
accuracy specificity, sensitivity, positive predictive value, and negative predictive value. Our experimental results show that
the accuracy, specificity and sensitivity of the method are 87.25%, 81.99% and 96.70%, respectively. The effectiveness of
the method in clinical segmentation was verified. Our study has demonstrated that cervical regional segmentation of col-
poscopic images based on HSV color space using K-means has high clinical utility and can help medical specialists in the
diagnosis of cervical cancer.

Keywords  Colposcopy image · Image segmentation · Image mirror reflection · HSV color space · K-means algorithm

Introduction process for cervical cancer, it is impossible to popularize

Cervical cancer is the second most common cancer among
women in the world, and it is also the most common cancer
in developing countries. According to a medical report based
on cancer [1], the survival rate of cervical cancer patients
is nearly 100% with early stage detection, but that rate
decreases to 20–50% in the advanced stage [2, 3]. Because
of the high complexity and high cost of the overall screening
Bing Bai and Pei-Zhong Liu has contributed equally to this work.
* Pei‑Zhong Liu
pzliu@hqu.edu.cn
1
College of Engineering, Huaqiao University,
Quanzhou 362021, China
2
College of Computer Science and Technology, Huaqiao
University, Xiamen 361021, China
3
Research Center of Apply Statistics and Big Data, Huaqiao
University, No. 668, Jimei Avenue, Xiamen 361021, China
more advanced screening techniques for cervical cancer in
resource-poor areas. Therefore, the incidence and mortality
of cervical cancer in the less-developed regions are still high
[4, 5]. According to the World Health Organization, there
are currently three main cervical cancer screening methods,
including human papillomavirus (HPV) testing, Papanico-
laou (Pap) smear test, and visual inspection with acetic acid
(VIA). The first two methods are relatively complex and
expensive, especially for patients in developing countries
[6]. Even in the first two approaches, colposcopy is still a
necessary procedure during biopsy detection. Therefore,
colposcopy is important and most suitable to be used as a
screening tool.
Colposcopy is a major diagnostic method for the detec-
tion of cervical lesions in women. The purpose of colpos-
copy is to identify and determine the severity of the lesion
so that a higher level of examination can be taken when
necessary. Colposcopy includes a visual assessment of the
lower genital (cervical, vulva, and vaginal) systems, and the

13
Vol.:(0123456789)

appearance of metaplastic epithelium consisting of transfor-
mation zones within the cervix region [7]. In order to obtain
more comprehensive image information from colposcopy, it
has a binocular microscope with an embedded white light
source. During the examination, the acetic acid solution
of 3–5% was applied to the cervical surface, resulting in
abnormal metaplasia epithelium becoming acetowhite (AW)
region. Cervical precancerous lesions and invasive cancers
show abnormal morphological features, which can be identi-
fied by colposcopy. However, lesion characteristics of cer-
vical epithelium, such as color characteristics and opacity,
edge division, and tissue shape, are only judged by a doctor
(a colposcopy specialist) to obtain a clinical diagnosis. Due
to the subjective nature of the examination, the accuracy of
colposcopy is highly dependent on the doctor’s experience
and expertise. Therefore, colposcopy has better sensitivity
and can avoid many biopsies [8].
The colposcopy image is a color image of the RGB
model, which contains important clinical features includ-
ing the cervical canal (CC), squamous epithelium (SE),
columnar epithelium (CE), and the transition zone (TZ)
between the latter two. This is the area of most cervical
cancers [9]. The imaging mode of the original colposcopy
image is relatively complex. As shown in Fig. 1, the origi-
nal colposcopic image contains the medical device and the
cervical region (the area surrounded by black curves). The
cervical region is the main investigation area for doctors’
diagnosis. It also contains areas of high light interference
caused by Acetowhite Epithelium original (AW) and spec-
ular reflections (SR) due to cervical mucus. The specular
Fig. 1  Manual segmentation based on SR, cervical region and AW of
the original colposcopic images by human experts. A, B, C, D repre-
sents colposcopy images of four patients, respectively, which have the
same structure.
13
Australasian Physical & Engineering Sciences in Medicine
reflection will lead to uneven imaging of cervical epithe-
lial tissue and may interfere with segmentation of cervi-
cal lesions [10]. Therefore, in the clinical diagnosis, the
original image of the colposcopic must first be subjected
to de-mirror reflex pretreatment. Then, the preprocessed
colposcopy image is segmented to achieve accurate deline-
ation of the cervical region (ROI), thereby providing a
regional guide for the diagnosis by medical experts.
In response to the above issues and the application
needs of clinical cervical cancer screening, Obukhova
et al. [11] proposed embedded artificial intelligence in
colposcopy to enable the colposcopy imaging device to
automatically analyze the results of the acetic acid test.
The key premise of this task is to automatically segment
the image of the cervical region and develop a set of cer-
vical segmentation (ROI) segmentation and segmentation
procedures. The device obtains colposcopy images under
different wavelengths of light and performs high-light
removal operation on the images. The diagnosis results are
obtained by cross-analysis of discrete Fourier transforms
of different spectrograms. In the research project of Patil
et al. [12], a special camera with a green filter was used
to collect colposcopic images. They removed high-light
areas by setting the pixels with higher color gradients as
specular areas and filling them with pixels. Finally, the
K-means was used to segment the colposcopic image in
the region of interest (ROI). Huang et al. [13] proposed a
transformational learning algorithm based on image pixel
classification that can also roughly segment the cervical
region. However, this method ignores the removal and fill-
ing of the specular reflection region formed during the
colposcopy imaging. Praba et al. [14] applied a Gaussian
mixture modeling method based on mathematical mor-
phology prior to feature extraction to eliminate irrelevant
information and specular reflection in colposcopic images.
In addition, sparse reconstruction-based image segmenta-
tion methods have also been applied to the segmentation
of the cervical region [15].
Based on the literature review, it can be seen that the use
of colposcopy for cervical cancer screening has two basic
problems of removing specular reflex and cervical seg-
mentation. Considering that the expert’s diagnosis process
requires more comprehensive information, the pretreatment
process of specular reflection removal needs to retain more
image textures and color features, and the segmentation pro-
cess of the cervical region needs to be efficient, accurate, and
complete. In this paper, based on the scaling of the original
colposcopy image, a specular reflection removal and fill-
ing algorithm based on image texture and color features is
proposed. The unsupervised clustering of cervical regions
based on the V component of the HSV color features are also
used to define the region of interest (ROI) as the segmenta-
tion method.
Australasian Physical & Engineering Sciences in Medicine
Colposcopic images differ from real cervical regions,
and doctors can use computer vision techniques to reduce
discrepancies for accurate diagnosis. This paper proposes a
segmentation method based on HSV color mode. The seg-
mentation accuracy of this algorithm reaches 82–97% under
the premise of removing highlights, the positive predictive
value (PPV) and negative predictive value (NPV) are 82.21%
and 96.80%. This method also greatly improved the clinical
diagnosis rate.
Methods
Algorithmic framework for cervical segmentation
The overall idea of the cervical region segmentation algo-
rithm in the colposcopic image based on the HSV color
mode is shown in Fig. 2. This process consists of two com-
ponents, including preprocessing to remove specular reflec-
tions and the clustering algorithm based on the V component
in the HSV color model. Specular reflection refers to non-
saturating bright pixels or patches that may occur during
medical imaging. This type of specular reflex is a common
cause of colposcopic and colposcopic image quality decline
[16]. Physicians usually prefer to eliminate high light reflec-
tance in medical images so that more useful clinical infor-
mation can be obtained [17]. In the case of colposcopy or
endoscopy, the flashlight carried by the imaging device pro-
duces specular reflections on the inner surface of the cervix.
Due to the large variation in the shape and color of different
images, the detection of specular reflection is an important
prerequisite for eliminating and filling the area.
As shown in the upper part of Fig. 2 the specular reflec-
tion section of the original colposcopy image is detected
according to the set threshold value, and then the average
luminance value of the neighboring region of the highlight
portion is selected for luminance replacement. Finally, the
saturation (S) patch is applied to this part using the average
saturation of the neighborhood. Then pre-processed image is
Fig. 2  Algorithmic framework of the colposcopic image processing
of the cervical region of interest
used as input image for segmentation of the cervical region.
The lower part of Fig. 2 shows that the V component in
the colposcopic image of the HSV color mode is extracted
according to a clustering method to segment the cervical
region. The cervix region in the colposcopic image occupies
the center of the image, so the cervix region segmentation
can be performed within a certain range from the center of
the image according to the similarity of the V component.
Figure 2 clearly shows the flow of the cervical region
segmentation method. The preprocessing part of this method
focuses on the luminance component (Y) of the colposcopic
image, which is also described in detail later in this section.
Similarly, after the K-means segmentation, the operation of
the image, the area filtering, the construction template, and
the like are required to finally obtain the segmented cervi-
cal region.
Removal of Specular Reflection
In the colposcopic image, the proportion of the specular
reflection area to the entire image is small, and the original
information of the epithelial tissue covered by the reflec-
tion area cannot be recovered. Therefore, the removal of
these highlight areas requires the restoration of the image
of the epithelial tissue in the area by the region filling algo-
rithm [18]. In the lesion area of the cell section, the lesion
cells in the CIN1 tissue were discontinuous or patchy, and
the positive intensity was weak. In the tissues of CIN2 and
CIN3, they are characterized by full-thickness continuous
and diffuse strong positives [19, 20]. The specular reflection
area usually has the characteristics of high brightness and
low color saturation [21]. To eliminate these specular reflec-
tions, the specular reflection area can be filled with a similar
image texture and color as the patch [22]. The following
content will introduce the relevant theoretical principles of
high light detection and removal of highlights and filling.
Colposcopic image analysis is often an analysis of the chro-
maticity component (U&V) and the luminance component
(Y) of the image, with highlight detection focusing more on
the luminance component (Y) [23].
Firstly, the image is converted from the RGB color mode
to the YUV color space using Eq. (1). Then, the highlight
area is detected according to the brightness threshold set
in Eq. (2), and Y(p) represents the pixel brightness value.
Firstly, the luminance histogram and the luminance surface
of the colposcopy image is calculated. The brightness of
the specular reflection area of each colposcopic image is
greater than gray scale 220, which can be determined by the
value of the overlap from the luminance histograms on the
different colposcopic images. Finally, the threshold involved
in removing the specular reflection is set to be between 220
and 255 in conjunction with the advice of the doctors (cer-
vical experts). This experiment is also facilitated by setting
13

reasonable brightness threshold. The high brightness area
in the colposcopic image shows low saturation. Therefore,
removing the high brightness area requires repairing the
saturation of the area. The area with specular reflection is
generally the smooth surface of the endoscope and cervix.
Finally, according to the Eq. (3), the saturation of the neigh-
borhood of the highlight region is selected and its saturation
is filled, where N8(p) is the saturation of the eight-neighbor-
hood region.
⎡Y ⎤ ⎡ 0.299 0.587 0.114 ⎤ ⎡R ⎤
⎢ ⎥ ⎢ ⎢ ⎥
⎢U ⎥ = ⎢ −0.147 −0.289 0.436 ⎥⎥ ⋅ ⎢G⎥ (1)
⎢V ⎥ ⎣ 0.615 −0.515 −0.100 ⎦ ⎢B ⎥
⎣ ⎦ ⎣ ⎦
{ change in the center of the two adjacent clusters, it means
Y(p) ≤ 255
det(p) = (2)
Y(p) ≥ 220
fill(p) = mean(N8(p)) (3)
The pre-processing of the colposcopic image was com-
pleted by the above highlight detection and saturation (S)
patching and was used as an input for segmentation of the
cervical region.
Cervical region segmentation algorithm (K‑means)
The colposcopic images of the preprocessed RGB color
mode were converted into HSV color space, and the V com-
ponent was extracted using the K-means algorithm. Finally,
the area filter is used to smooth the edges so as to segment
the cervical region. We then convert RGB color mode col-
poscopic image to HSV color space [24] as follows:
V = max(R, G, B) (4)
⎧ V − min(R, G, B)
⎪ , V≠0
S=⎨ V (5)
⎪0, otherwise
⎩ the remaining pi i=1 and assign them to the nearest
⎧60 ∗ (G − B)∕ (V − min(R, G, B)), V = R
⎪ {
H = ⎨120 + 60 ∗ (B − R)∕ (V − min(R, G, B)), V = G (6)
⎪240 + 60 ∗ (R − G)∕ (V − min(R, G, B)), V = B
⎩ Step 3: Update the point selected in the second step by
K-means algorithm is a fast and simple clustering
algorithm, which divides the objects into K classes [25].
The basic idea of the partitioning method is that given a
data set with N tuples or records, the split method will
construct K groups, each group represents a cluster,
and ensure that K < N. And these K packets satisfy the
13
Australasian Physical & Engineering Sciences in Medicine
following conditions: (1) Each packet contains at least
one data record. (2) Each data record belongs to one and
only one group. For a given K, an initial grouping method
is given first, and then the grouping is changed through
repeated iterative methods, so that the grouping scheme
after each improvement is better than the previous one.
A good standard is that the records in the same group
should be as close as possible, while the records in differ-
ent groups should be as far away as possible.
The K-means algorithm divides the pixels in the colpo-
scopic image into K cluster centers, then computes the dis-
tance from each sample to the cluster center, and classifies
the sample to the cluster center nearest to it. The algorithm
is as follows: Calculate the average value of each clustered
data object to getting a new cluster center. If there is no
that the sample adjustment is finished and the clustering
criterion
{ }function has converged. Hypothesis is set as
N,M
P = pi,j i=1,j=1 , where p is the pixel, N, M is the number of
( )
rows and columns; pi = hi , si , vi represents the { pixels
}K of
the HSV format after the color space conversion. p̂ j j=1 and
p̂ j = (hj , sj , vj ) indicate the pixels that need to be clustered in
the image. Equation (7) indicates the minimum Euclidean
distance between the calculated pixel points and the cluster
center in the K-means algorithm [26]. As the value of K
changes, the points that are similar to the cluster center are
gradually categorized into one.
∑
L
D(K) = min (pi − p̂ j )2 (7)
i=1
The K-means segmentation based on the HSV color
space characteristics of the colposcopic image is mainly
achieved through the following five steps:
( )
Step 1: Select K cluster centers from pi = hi , si , vi ran-
domly.
Step 2: Measure their distance to each cluster center for
{ }N,M
cluster center. This step can be calculated by Eq. (8) for
each point to each cluster center. Then get the shortest
distance.
}
C(a) = pi |∀a ≠ b, ||pi − p̂ a ||2 < ||pi − p̂ b ||2 , (8)
Eq. (9.)
1 ∑
C(a) = p (9)
|C(a)| i∈C(a) i
Step 4: The value of a gradually increases until all points
are traversed and the image is divided into clusters
Australasian Physical & Engineering Sciences in Medicine

formed by K cluster centers. Through the above four negative (TN) indicates the correct background point; (4) false
steps, the colposcopic image is divided into three cluster negative (FN), the wrong background point. Table 1 shows
centers: the cervical region, the extra cervical region, quantitative statistics of the segmentation accuracy of all test
and the endoscopic region. images. The Jaccard coefficient (JI) in the table is calculated
Step 5: Filter the area of the cervical region to achieve the according to Eq. (10), where P0 represents the cervical region
purpose of smoothing the edges of the image. We per- obtained by manual colposcopy, and P1 represents the cervix
form area filtering on the binary image of the colpos- region segmented by the algorithm [27]. The Dice coefficient
copy image to remove the connected region with an area is calculated from Eq. (11) and represents the degree of coin-
less than 2000 and using 8-neighbor filter. The detailed cidence between the two pictures [28].
algorithmic process is: (1) Determine the connected The accuracy (Acc), specificity (Spe), sensitivity (Sen),
components. (2) Compute the area of each component. positive predictive value (PPV) and negative predictive value
(3) Remove small objects. We suppress the noise of the (NPV) of the model can be calculated based on the above four
colposcopy image while preserving the image detail fea- pixel classification types. And they are used as an important
tures as much as possible. This method can effectively indicator to evaluate the performance of the segmentation
remove the portion of the colposcopy image that does method. ACC​means that the pixel with the correct pixel size
not belong to the cervical region. occupies all pixels; Spe means the ratio of the correct cervix
region to the total ROI region; and Sen refers to the ratio of the
Experimental preparation correct background point to the total background point. PPV
indicates the possibility that the segmentation point does not
A good experimental platform is important for perform- include the background point, and NPV indicates the value
ing the analysis of colposcopic images. The configuration of the possibility that the background point includes the seg-
of the computer is as follows: CPU: Inter Celeron(R) CPU mentation point [29–31]. The computed results are shown by
G550@2.60  GHz, Memory: 8  GB, Operating System: Eqs. (10, 11, 12, 13, 14, 15, 16).
Windows 64-bit. 110 colposcopic images were used as test
|P0 ∩ P1|
samples for the algorithm. The case images were provided JI = × 100% (10)
|P0 ∪ P1|
by Fujian Provincial Maternal and Child Health Hospital.
All the pictures in this study were provided by the Fujian
2|P0 ∩ P1|
Maternal and Child Health Hospital, which were provided Dic = × 100% (11)
by 100 patients, and they were from the same hospital. Half |P0| + |P1|
of 100 patients are cervical cancer patients. These images
are tagged by several doctors (experts) who are between the Acc =
TP + TN
× 100% (12)
ages of 40 and 60. The age of the patients is between 20 TP + FN + FP + TN
and 60 years old. The image format was 8-bit JPG color
image. The original image resolution is 6000 × 4000, in Sen =
TP
× 100% (13)
order to improve the efficiency of the algorithm and ensure TP + FN
the integrity of the image information, the image resolution
is adjusted to 6000 × 4000 in proportion. Each image was Spe =
TN
× 100% (14)
taken from a different patient. In order to match the medical TN + FP
expert’s initial diagnosis, all the samples were the original
colposcopic images taken before the treatment of acetic acid
and iodine solution. Table 1  Accuracy of the cervix region extraction compared with the
manual segmentation results
Performance indicators
Performance Maximum Minimum Median (%) Mean (%)
index stand- (%) (%)
The purpose of segmentation of the cervical region is to divide ard
the pixels of the colposcopy image into two ROI points and
JI 94.35 53.41 88.04 79.61
background points. In order to judge the segmentation result,
Dic 97.09 69.63 93.64 88.42
a reasonable evaluation index must be introduced. Compar-
Acc 87.38 87.15 87.22 87.25
ing cervical region segmentation results with expert annota-
Spe 98.07 53.43 90.20 81.99
tion results, the final pixels are divided into the following four
Sen 100 82.14 97.35 96.70
categories: (1) true positive (TP), which means the correct
PPV 80.41 80.03 80.17 80.21
segmentation of the cervical region; (2) false positives (FP)
NPV 96.86 96.71 96.82 96.80
indicates the point of excision of the cervical defect; (3) true

13

TP
PPV = × 100% (15)
TP + FP
TN
NPV = × 100% (16)
TN + FN
Experimental results
Image analysis results
The original colposcopic image format is RGB format.
Using different color space modes helps to segment the
cervical region in the colposcopic image. The HSV format
color feature can obtain more abundant clinical information
of the colposcopic image. The main feature of the highlight
region is color unsaturation, and the V (lightness) compo-
nent in the value HSV is saturated [32], as shown in Fig. 3.
According to the characteristics of colposcopy imaging, the
region covered by the V component in the HSV color space
is more easily extracted due to the texture characteristics of
the uterine cervix region, which is the cervical region that
needs to be segmented in clinical.
The K-means algorithm was used to process the origi-
nal colposcopic images after SR removal pre-processed,
and three cluster center clusters were obtained, including
the medical device, the outer region of the cervix, and the
cervical region. According to the colposcopy imaging char-
acteristics, the cervical region is located in the center of the
Fig. 3  The components of the colposcopic image, (A1, A2, A3) Original colposcopic image, (B1, B2, B3) HSV format image, (C1, C2, C3) H
component image, (D1, D2, D3) S component image, (E1, E2, E3) V component image
13
Australasian Physical & Engineering Sciences in Medicine
colposcopic image. Then, the cluster center cluster located in
the middle of the image is extracted, and the portion includes
the cervical region basically. Finally, the area filter is used
to remove the tiny connected areas in the cluster center and
smooth the edge of the cervical region.
Figure 4 shows an image of the cervical region segmented
from the de-coarsened speculum image using the method of
the present disclosure. The figure shows the segmentation
effect of four groups of cases, the third column represents the
segmented cervical region map. Clinically, the doctor will
make a diagnosis based on the cervical area map. It can be
seen that the segmented cervical region includes all regions
of interest basically in the standard data. Table 1 shows the
quantitative information of the segmentation results.
Performance indicators analysis
The quantitative descriptions of the segmentation algorithm
are detailed in Table 1. The JI and Dic coefficients in Table 1
are used to represent the segmentation results with the dis-
placement of the standard data. We hope that our judging
criteria (diagnostic criteria) are as high as possible for Sen,
Spe and Acc. Similarly, the PPV and NPV in Table 1 are as
high as possible, which means that the performance of the
algorithm is superior.
The experimental results show that the average JI and
Dic coefficients in the 110 test images, as marked by the
experts, are 79.61% and 88.42%, respectively. It can be seen
that the ratio of the intersection between the result image
Australasian Physical & Engineering Sciences in Medicine

Fig. 4  Cervical region segmentation results, (A1, A2, A3, A4) Colposcopy original image; (B1, B2, B3, B4) V component image; (C1, C2, C3,
C4) Cervical image; (D1, D2, D3, D4) Image after cervical region binarization operation; (E1, E2, E3, E4) Results image of expert marking

and the standard image is more than 80%. The results show
that the segmented cervical region and the expert-standard
ROI region have a high degree of coincidence in position.
The positive predictive value (PPV) and negative predictive
value (NPV) in the test results remained at 80% and 96%.
Therefore, the average sensitivity of the method is stable at
96.70%. The average sensitivity indicates that the method
can segment 96% of the cervical region in the colposcopy
image. The specificity represents the proportion of correctly
segmented image areas occupied by experts, and the average
specificity of this paper also reaches 82%. The Acc value can
best represent the segmentation accuracy of the method [33].
The Acc value of this method is stable at 87%, indicating that
this method can better meet the medical expert or specialist’s
diagnosis in clinical practice.
Discussion
The colposcopy image is the original image acquired by the
camera. In this experiment, the highlight area is removed
firstly, and then the cervical region is divided. We use three
methods: (1) the improved segmentation method of the
paper. (2) K-means segmentation algorithm based on LAB
color mode. (3) Maximum between-class variance (Otus
Method). The indicators of the three algorithms are pre-
sented in Table 2.
In this paper, we use the ground truth provided by doctors
to compare the segmentation accuracy of three segmentation
cervical region algorithms. The [bold] in Table 2 indicates
the best performing evaluation index in the corresponding
algorithm. In this experiment, the segmentation results of
the three algorithms are compared with the corresponding
ground truth. The K-means segmentation algorithm based on
LAB color mode is proposed by Das A et al., and the JI and
Dic of the algorithm are 77.56% and 82.31%, respectively
[34]. The displacement of the segmentation effect of the

Table 2  Results of three Contrast algorithm JI (%) Dic (%) Acc (%) Spe (%) Sen (%) PPV (%) NPV (%)
comparison algorithms
K-means (LAB) 77.56 82.31 83.45 80.19 90.03 82.45 89.84
OSTU 80.75 80.12 82.42 77.83 89.02 78.89 86.57
Our method 79.61 88.42 87.25 81.99 96.70 80.21 96.80

13

OSTU cervical segmentation algorithm proposed by Tra-
versi et al [35]. The higher the Acc, Spe and Sen, the better
the segmentation effect. The segmentation accuracy of the
algorithm is the highest, and the accuracy of segmentation
is 87.25%. The specificity of each algorithm is close to 80%,
and the sensitivity of our algorithm is 96.70%. The PPV
index of K-means (LAB) is higher than other algorithms,
indicating that the algorithm can segment the pixels belong-
ing to the cervical region as much as possible. The NPV of
the proposed algorithm is higher than the other two types
of algorithms, which indicates that the effect map obtained
by the method segmentation contains the least incoherent
regions.
Conclusion
In this paper, a new cervical segmentation method was
developed by using the K-means algorithm to segment the
colposcopic image in HSV color format. The technique first
uses the set brightness(Y) threshold to perform pre-process-
ing of highlight removal on colposcopic images. Then, the
V component in the input colposcopic image is extracted
to achieve the purpose of segmenting the cervical region.
The method was validated using 110 clinical data labeled
by medical experts. Sensitivity (Sen), specificity (Spe),
and accuracy (Acc) of the cervical segmentation method
are 96.70%, 81.99%, and 87.25%, respectively. Our new
approach has a better segmentation performance in com-
parison to the traditional algorithm, and can be used poten-
tially as an image analysis platform to promote the clinical
prognosis of cervical cancer. The next work outlook is to (1)
improve the accuracy of the algorithm for segmentation of
the cervical region. Can really apply to clinical medicine.
(2) Based on the cervical region, the cervical lesions were
extracted. Combining reality, conducting in-depth research.
Acknowledgements  We thank the Fujian Provincial Maternal and
Child Health Hospital for providing the datasets used in this paper.
Funding  This work was supported by the Grants from National Natu-
ral Science Foundation of China (No. 61605048 and No. 61603144,
and Grant 61403245 and Grant 91648119), Natural Science Founda-
tion of Fujian Province, China (No. 2016J01300), Promotion Pro-
gram for Young and Middle-aged Teacher in Science and Technology
Research of Huaqiao University (No. ZQN-PY518), and the Scientific
Research Funds of Huaqiao University (No. 15BS413), and the Young
and Middle-aged Teachers Education Scientific Research Project of
Fujian Province, China (No. JAT160020), Natural Science Foundation
of Fujian Province, China (Grant No. 2015J01256), the Talent project
of Huaqiao University (Grant No. 14BS215), and Quanzhou scien-
tific and technological planning projects of Fujian, China (Grant Nos.
2015Z120, 2017G024), and the Subsidized Project for Postgraduates’
Innovative Fund in Scientific Research of Huaqiao University (No.
17014084001).
13
Australasian Physical & Engineering Sciences in Medicine
Compliance with ethical standards 
Conflict of interest  The authors declare that there are no conflicts of
interest.
Ethical approval  All experiments were performed in compliance with
the ethical standards set by our institutional board.
Informed consent  Informed consent was obtained from all individual
participants included in the study.
References
1. Chen W, Zheng R, Baade PD, Zhang S, Zeng H, Bray F et al
(2016) Cancer statistics in china 2015. CA Cancer J Clin
66(2):115–132
2. Pfaendler KS, Wenzel L, Mechanic MB, Penner KR (2015) Cervi-
cal cancer survivorship: long-term quality of life and social sup-
port. Clin Therapeut 37:39–48
3. Dempsey C, Govindarajulu G, Sridharan S et al (2014) Implica-
tions for dosimetric changes when introducing MR-guided brachy-
therapy for small volume cervix cancer: a comparison of CT and
MR-based treatments in a single centre. Australas Phys Eng Sci
Med 37(4):705–712
4. Cronje HS (2004) Screening for cervical cancer in developing
countries. Int J Gynecol Obstet 84:101–108
5. Sun YP, Sargent D, Lieberman R, Gustafsson U (2011) Domain-
specific image analysis for cervical neoplasia detection based on
conditional random fields. IEEE Trans Med Imaging 30:867–878
6. Sun YP, Sargent D, Lieberman R et al (2011) Domain-specific
image analysis for cervical neoplasia detection based on condi-
tional random fields. IEEE Trans Med Imaging 30(3):867–878
7. Gordon S, Zimmerman G, Greenspan H (2004) Image segmenta-
tion of uterine cervix images for indexing in PACS. In: Proceed-
ings of the IEEE symposium on computer-based medical systems,
pp 298–298
8. Li W, Poirson A (2006) Detection and characterization of abnor-
mal vascular patterns in automated cervical image analysis,
Advances in Visual Computing. Springer, Berlin, pp 627–636
9. Van Raad V (2003) Design of Gabor wavelets for analysis of tex-
ture features in cervical images. Engineering in Medicine and
Biology Society. In: Proceedings of the international conference
of the IEEE, pp 806–809
10. Liu J, Li L, Wang L (2018) Acetowhite region segmentation in
uterine cervix images using a registered ratio image. Comput Biol
Med 93:47–55
11. Obukhova NA, Motyko AA, Kang U et al (2017) Automated
image analysis in multispectral system for cervical cancer
diagnostic. In: Conference of Open Innovations Association,
pp 345–351
12. Patil DB, Gaikwad MS, Singh DK et al (2016) Semi-automated
lession grading in cervix images with Specular Reflection
removal. In: International conference on inventive computation
technologies, pp 1–5
13. Huang S, Gao M, Yang D et al (2015) Unbalanced graph-based
transduction on superpixels for automatic cervigram image seg-
mentation. In: IEEE international symposium on biomedical
imaging, IEEE, pp 1556–1559
14. Praba PSR, Ranganathan H (2013) Wavelet transform based auto-
matic lesion detection in cervix images using active contour. J
Comput Sci 9(1):30–36
Australasian Physical & Engineering Sciences in Medicine
15. Wang W (2010) Cervigram image segmentation based on recon-
structive sparse representations. Med Imaging Image Process.
https​://doi.org/10.1117/12.84546​1
16. Xue Z, Antani S, Long LR et  al (2007) Comparative perfor-
mance analysis of cervix ROI extraction and specular reflection
removal algorithms for uterine cervix image analysis. Proc SPIE
6512(1):187–189
17. Karapetyan G, Sarukhanyan H (2013) Automatic detection and
concealment of specular reflections for endoscopic images. In:
Computer science and information technologies, IEEE, pp. 1–8
18. Guillemot C, Meur OL (2013) Image inpainting: overview and
recent advances. IEEE Signal Process Mag 31(1):127–144
19. Castle PE, Stoler MH, Solomon D et al (2007) The relationship
of community biopsy-diagnosed cervical intraepithelial neoplasia
grade 2 to the quality control pathology-reviewed diagnoses: an
ALTS report. Am J Clin Pathol 127(5):805–815
20. Corrêa FM, Russomano FB, Oliveira CA (2012) Colposcopic
triage methods for detecting cervical intraepithelial neoplasia
grade 3 after cytopathological diagnosis of low-grade squamous
intraepithelial lesion: a systematic review on diagnostic tests. São
Paulo Med J Revista paulista de medicina 130(1):44–52
21. Xue Z, Antani S, Long LR et al (2007) Comparative performance
analysis of cervix ROI extraction and specular reflection removal
algorithms for uterine cervix image analysis. Medical Imaging
2007: Image Processing. International Society for Optics and Pho-
tonics, pp 187–189
22. Maheswari GU, Ramar K, Manimegalai D et al (2011) Short
communication: an adaptive region based color texture seg-
mentation using fuzzified distance metric. Appl Soft Comput J
11(2):2916–2924
23. Stehle TH (2006) Specular reflection removal in endoscopic
images. In: Proceedings of the 10th international student confer-
ence on electrical engineering, vol 10, pp 1–6
24. Mohd Ali N, Md Rashid NKA, Mustafah YM (2016) Performance
comparison between RGB and HSV color segmentations for road
signs detection. Appl Mech Mater 393:550–555
25. Luo M, Ma YF, Zhang HJ (2003) A spatial constrained K-means
approach to image segmentation. Joint Conference of the Fourth
International Conference on Information, Communications and
Signal Processing, 2003 and Fourth Pacific Rim Conference on
Multimedia. IEEE, vol 2, pp 738–742
26. Siddiqui FU, Isa NAM (2011) Enhanced moving K-means
(EMKM) algorithm for image segmentation. IEEE Trans Consum
Electron 57(2):833–841
27. Real R, Vargas JM (1996) The probabilistic basis of Jaccard’s
Index of similarity. Syst Biol 45(3):380–385
28. Dice LR (1945) Measures of the amount of ecologic association
between species. Ecology 26(3):297–302
29. Gerig G, Jomier M, Chakos M (2001) Valmet: a new valida-
tion tool for assessing and improving 3D object segmentation.
In: International conference on medical image computing and
computer-assisted intervention. Springer, pp 516–523
30. Udupa JK, Leblanc VR, Zhuge Y, Imielinska C, Schmidt H, Cur-
rie LM, Hirsch BE, Woodburn J (2006) A framework for evaluat-
ing image segmentation algorithms. Comput Med Imaging Graph
30(2):75–87
31. Panigrahy D, Sahu PK (2017) Extraction of fetal ECG signal by
an improved method using extended Kalman smoother framework
from single channel abdominal ECG signal. Australas Phys Eng
Sci Med 40(1):191–207
32. Saint-Pierre CA, Boisvert J, Grimard G, Cheriet F (2011) Detec-
tion and correction of specular reflections for automatic surgical
tool segmentation in thoracoscopic images. Mach Vision Appl
22(1):171–180
33. Proença H, Alexandre LA (2010) Iris recognition: analysis of
the error rates regarding the accuracy of the segmentation stage.
Image Vision Comput 28(1):202–206
34. Das A, Kar A, Bhattacharyya D (2011) Elimination of specular
reflection and identification of ROI: the first step in automated
detection of Cervical Cancer using Digital Colposcopy. IEEE Int
Conf Imaging Syst Techn 5(3), 237–241
35. Traversi M, Falagario M, Guaragnella C (2014) CADdy—Col-
poscopy learning machine for computer aided diagnosis. In: IEEE
third international conference on consumer electronics, Berlin,
IEEE, vol 7(6), pp 1–4
13