This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

Automatic classification of cervical cells using deep

learning method
Suxiang Yu1, Xinxing Feng2, Bin Wang1, Hua Dun1, Shuai Zhang3, Ruihong Zhang4, and Xin
Huang5
1
Department of Pathology, The Fourth Central Hospital of Baoding City, Hebei,, 072350, People's Republic of China
2
Endocrinology and Cardiovascular Disease Centre, Fuwai Hospital, National Center for Cardiovascular Diseases, Chinese Academy of Medical
Sciences and Peking Union Medical College, Beijing, 100037, People's Republic of China
3
Department of Computer Science, The University of Manchester, Manchester, M13 9PL, UK
4
Department of Science and Teaching, The Fourth Central Hospital of Baoding City, Hebei, 072350 People's Republic of China
5
Solar Activity Prediction Center, National Astronomical Observatories, Chinese Academy of Sciences, Beijing 100012, People’s Republic of China

Co-first authors: Xinxing Feng

Co-corresponding author: Ruihong Zhang (e-mail: [email protected])
Corresponding author: Xin Huang (e-mail: xhuang@ bao.ac.cn).

ABSTRACT

Cervical cancer is the fourth most prevalent disease among women. Prompt diagnosis and its management
can significantly improve patients’ survival rates. Therefore, routine screening for cervical cancer is of
paramount importance. Herein, we explore the potential of a deep learning model to automatically distinguish
abnormal cells from normal cells. The ThinPrep cytologic test dataset was collected from the fourth central
hospital of Baoding city, China. Based on the dataset, four classification models were developed. The first
model was a 10-layer convolutional neural network (CNN). The second model was an advancement of the
first model equipped with a spatial pyramid pooling (SPP) layer (CNN + SPP) to treat cell images based on
their sizes. Based on the first model, the third model replaced the CNN layers with the inception module
(CNN + Inception). However, the fourth model incorporated both the SPP layer and the inception module
into the first model (CNN + inception + SPP). The performances of the four models are estimated and
compared by using the same testing data and evaluation index. The testing results demonstrated that the fourth
model yields the best performance. Moreover, the area under the curve (AUC) for module four was 0.997.

INDEX TERMS: Cell classification, Deep learning, Neural networks, Cervical cytology

I. INTRODUCTION processing and machine learning technologies, computer-

Cervical cancer refers to some cervix cells rapidly assisted cervical screening methods are proposed to examine
becoming malignant [1]. Besides, it is among the prime the cells in the whole image and categorizes them as normal
reasons for cancer death in women [2]. However, cervical and abnormal [3].
cancer can be effectively treated at an early stage. Therefore, The core of the automatic screening system is to develop
routine screening for cervical cancer is of paramount classification models using the machine learning method [4].
importance. Traditional machine learning manually extracts features from
The screening process involves either Papanicolaou smear the images of cells before establishing a relationship between
or ThinPrep cytologic tests (TCTs) and subsequent features and classes. Besides, hand-crafted features limit the
examination under a microscope by a pathologist for the performance of the model. Nevertheless, deep learning [5,6]
presence of abnormal cells. There are thousands of cells per integrates feature extraction and classification into one
the testing result, hence the pathologist carefully scans and optimized process. There is no need to design hand-crafted
makes a judgment. This is a time-consuming process with features because it can learn effective features and yield
subjective or biased experiences. With advances in image satisfactory performance.

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

Deep learning requires a large number of labeled cell Manually distinguishing abnormal cells from normal cells
images with fixed size to develop a classification model. It is is time-consuming, subjective to bias and prone to errors.
time-consuming to label the cervical data because the size of Therefore, machine learning can be applied to develop an
cervical cells is not fixed. automatic classification model to improve the screening
Herein, we adopt the inception module [7, 8] following its system performance. A machine learning system should
potential in increasing the depth and width of the network with include data acquisition, feature extraction and modeling
a slight increase in the number of network parameters. method. Related works are summarized in Table I.
Elsewhere, the spatial pyramid pooling (SPP) layer [9] was Based on the dataset, the Herlev pap smear dataset was the
used to treat images with different cell sizes. frequently used public data [12-17,32]. Notably, UCI
Four deep learning models (convolutional neural network (digitized cervical cancer data set) was among the popular
(CNN), a CNN with inception module, a CNN with SPP layer, datasets in the cervical cancer dataset. However, the original
a CNN with inception module and SPP layer) were developed images were excluded, only extracted features were included
on the training set with both their estimated performances in these datasets [34,35]. Apart from public data, some private
compared during the testing set (Figure 1). Finally, the best datasets were also collected. For example, pap smear dataset
model was selected to distinguish abnormal cells from normal was collected from Fortis Hospital Mohali, India [27], liquid-
cells. based cytology datasets from People’s Hospital of Nanshan

Images of cervical
CNN: Convolutional Neural Network cells with labels
SPP : Spatial Pyramid Pooling

Training set Testing set

CNN+Inception+SPP
CNN CNN+Inception CNN+SPP

Performance Performance Performance Performance

evaluation evaluation evaluation evaluation

Performance comparison

District, China [10] and Alzahra Hospital of Tabriz, Iran [18]

FIGURE 1. THE FLOW CHART OF THE PROPOSED MODELS.
and pathological images stained with hematoxylin and eosin
To the best of our knowledge, this is the first work to adopt from Tianjin Tumor Hospital, China [19] and first Affiliated
a model with both the inception module and SPP layer to Hospital of Xinjiang Medical University, China [29].
automatically classify cervical cells. Regarding features, morphometric [11, 27, 28, 31, 32, 33],
The structure of this paper is as follows: textual [11, 19, 21, 23, 31, 32, 33], color histogram [19, 21,
Sec. II includes related works on automatic classification of 24], geometric features [26, 32] as well as local binary patterns
the cervical cell. Sec. III specifies the modeling methods. Sec. [16] were extracted from the nucleus and cytoplasm. Features
IV has experiments and results. Sec. V provides discussions of are not only extracted from a single cell but also from blocks
our model. Finally, Sec. VI has a brief conclusion of the work. of cells in whole slide cervical cell images to reduce the
computational complexity [19]. Although many features can
be extracted, not all are useful. Feature selection methods such
II. RELATED WORKS
as particle swarm optimization-based feature selection [14]

VOLUME XX, 2017 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

and genetic algorithm-based feature selection [15] are used to with hand-crafted features whose discriminative ability is
discard redundant features. limited. However, deep learning potentially comprehends
With modeling algorithms, several machine learning features from input images before integrating feature
algorithms have been adopted in developing classification extraction and model development into a unified framework.
models for cervical cell, for example, linear discriminant CNN [20, 22, 25, 29], a popular deep learning method in
analysis [12], K-nearest neighbor [12, 17, 27], neural networks treating images, has been applied in distinguishing abnormal
[10, 12, 13, 18], Bayesian classifier [12] and support vector cervical cells from normal cells.
machine [10, 11, 12, 16, 19]. To improve the single In this paper, we improve the automatic cervical cell
classification model performance, ensemble learning that classification model performance by incorporating both the
incorporates multiple models to solve problems was used in inception module and the SPP layer into the CNN.
developing the classification model [10, 14, 15, 21]. Incorporating the inception module facilitates faster training
In traditional machine learning methods, feature extraction in the classification model with fewer data and higher
and modeling algorithm are two separate processes. performance. However, the SPP layer enhances the processing
Classification model performances are difficult to improve of input images of any size in the classification model.

TABLE I
RELATED WORKS OF AUTOMATIC CERVICAL CELL CLASSIFICATION
Num. Data Features Methods Ref.
1 Liquid-based cytology data collected from Shape, size and texture of the nuclei Neural networks, Ensemble [10]
People’s Hospital of Nanshan District, China learning, Support vector
machine, Random forest
2 Pap smears dataset Morphometric and textual features of the cells Support vector machine [11]
3 3 Pap smear dataset Features calculated from nucleus and cytoplasm Bayesian classifier, Linear [12]
(ERUDIT, LCH, Herlev) discriminant analysis, K-nearest
neighbor, Neural networks,
Support vector machine
4 Herlev Pap smear dataset 20 features for nucleus and cytoplasm Neural network [13]
5 Herlev Pap smear dataset 20 features for nucleus and cytoplasm Ensemble learning [14]
6 Herlev Pap smear dataset 20 features for nucleus and cytoplasm Ensemble learning [15]
7 Herlev Pap smear dataset Local binary patterns Support vector machine [16]
8 Herlev Pap smear dataset Discriminative feature extraction method K nearest neighbor [17]
9 Liquid-based cytology dataset collected from Linear plot (linear transformation on specially Neural network [18]
ALZAHRAHOSPITAL of Tabriz, Iran customized axis)
10 Pathological images stained with hematoxylin Texture and color histogram features extracted Support vector machine [19]
and eosin collected from TianjinTumor from blocks of cells in whole slide cervical cell
Hospital, China image
11 Herlev Pap smear dataset and liquid-based Deep leaning features Convolutional neural networks [20]
cytology datasets
12 2 pap smear datasets (SIPaKMeD and Herlev Texture, shape and color features extracted Ensemble learning [21]
datasets) from the regions of segmented nuclei and
cytoplasm
13 Herlev Pap Smear dataset. Deep learning features Convolutional neural networks [22]
14 Histology images of hematoxylin and Texture, Cellular Features, Nuclei Features Ensemble learning [23]
eosinophil
15 6 histopathological image datasets (AQP1, Color, texture and deep learning features Multilayer hidden conditional [24]
AQP2, HIF1, HIF2, VEGF1 and VEGF2 random fields
datasets)
16 Herlev Pap Smear dataset. Deep learning features Convolutional neural networks [25]
and extreme learning machine
17 Pap smear dataset Geometric and texture features Support vector machine [26]
18 Pap smear dataset collected from Fortis Hospital Morphological features K nearest neighbor [27]
Mohali, India
19 Pap smear datasets (DTU dataset and Herlev 29 morphology features fuzzy c-means algorithm to [28]
dataset) classify
20 Pathological images stained with hematoxylin Deep learning features Convolutional neural network [29]
and eosin collected from the First Affiliated
Hospital of Xinjiang Medical University, China
21 Liquid-based cytology slides 28 features including 20morphological and C4.5 and LogicalRegression [31]
8texture features
22 3 Pap smear dataset 29 features including 20 morphological Trainable Weka Segmentation, [32]
(Herlev, Norup, MRRH) features, 3 geometric features and 6 texture sequential elimination, simulated
features annealing, fuzzy C‑means
23 Pap smear dataset collected from Rajah 44 features including 14 texture features and 30 SVM [33]
Muthiah Medical College, Annamalainagar shape features

VOLUME XX, 2017 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

24 UCI (digitized cervical cancer data set) 30 features such as age, smokes, HIV, HPV, Softmax classification with [34]
number of sexual partners stacked autoencoder model
25 UCI (digitized cervical cancer data set) 30 features such as age, smokes, HIV, HPV, Random forest [35]
number of sexual partners

model will have a fixed size of result without a limitation of

III. METHOD input size. The structure of the SPP layer is shown in Figure 3.

A. DATASET
The dataset of 2504 cell samples (1202 abnormal and 1302
normal samples) were collected from the fourth central
hospital of Baoding city, China. Pathologists scanned whole
images of TCT using a digital camera mounted on a
microscope and manually cutout single cells and categorized
them as normal and abnormal. Each cell was double-checked
by pathologists. The dataset was categorized into the training
set and the testing set as ratio of 8:2.
Generalization of the final model was ascertained using the
5-fold cross-validation.

B. MODELING METHOD
Here, we developed a method to classify normal and
abnormal cells using four related models. Model A was a basic
model and both Model B and C were improved based on
model A, respectively. Model D combined the advantages of
the aforementioned models.
Model A was a simple CNN. This model had different parts
such as convolutional (conv), non-linearity, pooling (pool) FIGURE 3. The structure of SPP.
layers, and fully connected (FC) layers. The input of Model A The structure of Model B incorporated with SPP before the
was a 128p x 128p image. The structure of Model A is shown FC layer (Figure 4).
in Figure 2. There are pooling layers after the 4th, 7th and 10th
convolutional layer.

FIGURE 2. The structure of Model A FIGURE4. The structure of Model B

To avoid the impact of the variable size for cell images, To improve performance of Model A, Model C replaced
Model B adds the SPP layer before the FC layers that made CNN layers with the inception module.
the model adaptable to different input sizes. Based on previous studies, the inception module, as a
Unlike the traditional pooling layer, SPP [9] used a fixed network, contains several branches that increase the width of
number of pooling windows to generate feature maps. By the network [7,8]. The structure of the inception module is
automatically adjusting the size and pooling stride of pooling shown in Figure 5. The inception module adds some branches
windows, the size of pooling was common on different sizes that differ from basic layers and combines them. This strategy
of input. Therefore, by combining every pooling result, the

VOLUME XX, 2017 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

can improve the fitting ability of models. The structure of Both Model A and C are designed to deal with the fixed size
Model C is displayed in Figure 6. without SPP. Following variations on image sizes in our
dataset from 951p x 580p to 41p x 34p, we elected the 128p x
128p as the formal input size.

IV. RESULTS
This work summarizes model performance based on
precision, sensitivity, specificity, accuracy, F1 score and area
under the curve (AUC).
Precision and sensitivity showed the exactness and
completeness of classifiers, respectively. Besides, specificity
showed the potential of a classifier to correctly classify normal
data. Accuracy showed that a classifier can correctly
categorize the two-class task. Considering both recall and
precision, the F1 score and AUC were defined to evaluate
performances of the classification model. The equations of
FIGURE 5. The structure of the inception module performance indices are as follows.
𝑇𝑃
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = (1)
𝑇𝑃+𝐹𝑃
𝑇𝑃
𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = (2)
𝑇𝑃+𝐹𝑁
𝑇𝑁
𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = (3)
𝑇𝑁+𝐹𝑃
𝑇𝑃+𝑇𝑁
𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (4)
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁
2∗𝑇𝑃
𝐹1𝑠𝑐𝑜𝑟𝑒 = (5)
2∗𝑇𝑃+𝐹𝑃+𝐹𝑁
2∗𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦∗𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦
𝐻𝑚𝑒𝑎𝑛 = (6)
𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦+𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦

A. TRAINING RESULTS
We programmed the algorithm in Python to conduct the
experiments using NVIDIA GeForce RTX 2070 8G,
Windows operating system, Intel® CoreTM i7-9700K CPU @
3.60GHz and 16 GB RAM. Additionally, we used PyTorch-
1.2.0 to develop a basic framework for the classification model
FIGURE 6. The structure of Model C of cervical cells.
Briefly, Model D combined SPP and the inception module NNI (Neural Network Intelligence), an open-source
to develop a more flexible model with high accuracy. The AutoML toolkit from Microsoft, was used to determine the
structure of Model D is shown in Figure 7. best combination of hyperparameters (including learning
momentum, learning rate and dropout rate). By using the SPP
Module for the model to manage different input sizes, the
batch size can only be 1.
The NNI results are shown in Figure 8. These results are
based on Model D to identify the best combination of the
learning rate, learning momentum and dropout rate.
MetisTuner was used as an optimization strategy. From the
NNI final results, the best combination of the aforementioned
hyperparameters had a learning momentum, learning rate and
dropout rate of 0.9, 0.0001 and 0.5, respectively.
We trained Model A for 400 epochs using a learning
momentum, learning rate and a batch size of 0.9, 0.0001and 1,
respectively. However, Models B, C and D were trained for
450, 600 and 600 epochs, respectively. We use random
flipping online to augment the training set.
Figure 9 shows the training results at every phase and the
FIGURE 7. The structure of Model D training accuracy and loss in Model A-D.
The receiver operating characteristic (ROC) curves of
C. DATA PREPROCESSING Model A-D are shown in Figure10. The inception module and

VOLUME XX, 2017 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

the SPP significantly improved the performance of the model. 98.28 ± 0.21%, 98.20 ± 0.23%, 98.25 ±0 .21% and 0.9971 ±
Details of test results are shown in Table II. Models C and B 0.000001, respectively. These results indicated that Model D
achieved similar performance. The performances of Models B is stable.
and C were better than that of Model A, and Model D achieved
the best performance (Table II). These results demonstrate that To compare the performance of our model with related
both the inception module and the SPP can improve the works, we used the popular Papanicolaou smear dataset-
performances for the classification of cervical cells. Herlev dataset to test the performance of the proposed Model
Furthermore, a combination of the inception module and the D. The comparison is given in Table IV. We achieved an AUC
SPP can achieve the best performance. of 0.975, which is better than ENS [16], Dis(S+M) [17],
We used 5-fold cross-validation to test the generalization of Ensemble [36] and RegressionConstraint[37] and close to
Model D. The testing results are shown in Table III. AlexN-3C, GoogLeNet-3C, ResNet-3C and DenseNet-3C,
Consequently, Model D resulted (5-fold cross-validation) in a respectively [38]. Hence, our model performed well on the
precision, sensitivity, specificity, accuracy, F1, H-mean and an famous cervical cancer classification dataset-Herlev dataset.
AUC score of 98.76 ± 1.10%, 97.67 ± 1.01%, 98.85 ± 0.94%,

FIGURE 8. Hyper-parameter of NNI results

TABLE II
TESTING RESULTS OF FOUR MODELS
Model precision (%) sensitivity (%) specificity (%) accuracy (%) F1 score (%) H-mean (%) AUC
A 87.6 97.1 87.4 92.0 92.1 92.0 0.953
B 93.5 95.9 93.9 94.8 94.7 94.8 0.967
C 93.4 94.6 93.9 94.2 94.0 94.2 0.966
D 97.5 98.3 97.7 98.0 97.9 98.0 0.997

TABLE III
5-FOLD CROSS-VALIDATION OF MODEL D
Fold precision (%) sensitivity (%) specificity (%) accuracy (%) F1 score (%) H-mean (%) AUC
1 97.53 98.34 97.70 98.01 97.93 98.02 0.9969
2 100.00 96.27 100.00 98.21 98.10 98.10 0.9966
3 97.91 97.10 98.08 97.61 97.50 97.59 0.9957
4 98.35 99.17 98.47 98.81 98.76 98.82 0.9970
5 100.00 97.48 100.00 98.79 98.72 98.72 0.9991
Final 98.76 ± 1.10 97.67 ± 1.01 98.85 ± 0.94 98.28 ± 0.21 98.20 ± 0.23 98.25 ± 0.21 0.9971 ± 0.000001

TABLE IV
PERFORMANCE COMPARISON OF DIFFERENT CERVICAL CELL CLASSIFICATION MODELS
Model AUC
ENS [16] 0.884
Dis(S+M) [17] 0.964

VOLUME XX, 2017 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

Ensemble [36] 0.97

RegressionConstraint [37] 0.94
AlexN-3C [38] 0.962±0.008
GoogLeNet-3C [38] 0.979±0.005
ResNet-3C [38] 0.978±0.018
DenseNet-3C [38] 0.970±0.013
The proposed model 0.975

FIGURE 9: Training accuracy and loss of Model A-D

VOLUME XX, 2017 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

FIGURE 10 Testing ROC curves of Model A-D

normal cells. By comparing the performances of these

V. DISCUSSION models, the best model (CNN + SPP + Inception) was
The original image and the heat map from Model D are determined with an AUC value of 0.997. This model can
shown in Figure 11. In the heat map, the highlighted parts input cell images with arbitrary size. From the analysis of
are considered areas of importance in classification. Notably, feature maps, the best model extracts features from the
the deep learning model extract features are mostly from the nucleus, cytoplasm and their boundary, respectively. These
cell nucleus and partly from the cytoplasm. Coincidentally, results were consistent with the understanding of
these parts are also crucial in traditional feature extraction pathologists. Recently most models are developed to treat
methods [11, 19, 21]. This implies that features extracted by cell images from the Papanicolaou smear test, we learned
the deep learning method are reasonable and reliable to the classification mode from cell images of TCT. It is
some degree. noteworthy that the TCT is more sensitive than the
Papanicolaou smear test in screening of cervical cancer. The
performance of our model was tested by using images of the
Papanicolaou smear test. We found that the proposed model
was effective not only for TCT, but also for the
Papanicolaou smear test.
The volume of data, especially the abnormal cells, was
limited. Thus, adequate data should be collected and labeled
or the image generation technology should be used to
improve the classification model performance in the future.
FIGURE 11: Original image and the heat map of Model D
ACKNOWLEDGMENT
The authors would like to thank anonymous reviewers
VII. CONCLUSION for their valuable suggestions.
In this paper, four deep learning models (CNN, CNN +
SPP, CNN + Inception, CNN + SPP + Inception) were built
to automatically distinguish abnormal cervical cells from

VOLUME XX, 2017 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

REFERENCES [20] Zhang L, Lu L, Nogues I, et al. DeepPap: deep convolutional

[1] "Defining Cancer". National Cancer Institute. 17 September 2007. networks for cervical cell classification[J]. IEEE Journal of
Retrieved 28 March 2018. biomedical and health informatics, 2017, 21(6): 1633-1643.
[2] F.Bray, J. Ferlay, I.Soerjomataram, R.L.Siegel, L.A.Torre, [21] Win K P, Kitjaidure Y, Hamamoto K, et al. Computer-Assisted
and.Jemal, ‘‘Global cancer statistics 2018: GLOBOCAN estimates Screening for Cervical Cancer Using Digital Image Processing of
of incidence and mortality worldwide for 36 cancers in 185 Pap Smear Images[J]. Applied Sciences, 2020, 10(5): 1800.
countries,’’ CA, Cancer J. Clin., vol. 68, no. 6, pp. 394–424, 2018. [22] Allehaibi K H S, Nugroho L E, Lazuardi L, et al. Segmentation and
[3] H. C. Kitchener, R. Blanks, G. Dunn, L. Gunn, M. Desai, R. Albrow, classification of cervical cells using deep learning[J]. IEEE Access,
J. Mather, D. N. Rana, H. Cubie, C. Moore, R. Legood, A. Gray, 2019, 7: 116925-116941.
and S. Moss, “Automation-assisted versus a manual reading of [23] Guo P, Banerjee K, Stanley R J, et al. Nuclei-based features for
cervical cytology (MAVARIC): a randomised controlled trial,” uterine cervical cancer histology image analysis with fusion-based
Lancet Oncol., vol. 12, no. 1, pp. 56–64, 2011. classification[J]. IEEE Journal of biomedical and health informatics,
[4] Li, C., Chen, H., Li, X. et al. A review for cervical histopathology 2015, 20(6): 1595-1607.
image analysis using machine vision approaches. ArtifIntell Rev [24] Li C, Chen H, Zhang L, et al. Cervical histopathology image
(2020). https://doi.org/10.1007/s10462-020-09808-7 classification using multilayer hidden conditional random fields and
[5] Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification weakly supervised learning[J]. Ieee Access, 2019, 7: 90378-90397.
with deep convolutional neural networks[C]//Advances in neural [25] Ghoneim A, Muhammad G, Hossain M S. Cervical cancer
information processing systems. 2012: 1097-1105. classification using convolutional neural networks and extreme
[6] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. learning machines[J]. Future Generation Computer Systems, 2020,
521, no. 7553, pp. 436–444, 2015. 102: 643-649.
[7] Szegedy C, Liu W, Jia Y, et al. Going deeper with [26] Kashyap D, Somani A, Shekhar J, et al. Cervical cancer detection
convolutions[C]//Proceedings of the IEEE conference on computer and classification using Independent Level sets and multi
vision and pattern recognition. 2015: 1-9. SVMs[C]//2016 39th international conference on
[8] Szegedy C, Ioffe S, Vanhoucke V, et al. Inception-v4, inception- telecommunications and signal processing (TSP). IEEE, 2016: 523-
resnet and the impact of residual connections on 528.
learning[C]//Thirty-First AAAI Conference on Artificial [27] Sharma M, Singh S K, Agrawal P, et al. Classification of the clinical
Intelligence. 2017. dataset of cervical cancer using KNN[J]. Indian Journal of Science
[9] He K, Zhang X, Ren S, et al. Spatial pyramid pooling in deep and Technology, 2016, 9(28): 1-5.
convolutional networks for visual recognition[J]. IEEE transactions [28] William W, Ware A, Basaza-Ejiri AH, et al. Cervical cancer
on pattern analysis and machine intelligence, 2015, 37(9): 1904- classification from Pap-smears using an enhanced fuzzy C-means
1916. algorithm[J]. Informatics in Medicine Unlocked, 2019, 14: 23-33.
[10] L. Zhang, H. Kong, C. T. Chin, S. Liu, X. Fan, T. Wang, and S. [29] Wu M, Yan C, Liu H, et al. Automatic classification of cervical
Chen, “Automation-assisted cervical cancer screening in manual cancer from cytological images by using convolutional neural
liquid-based cytology with hematoxylin and eosin staining,” Cytom. network[J]. Bioscience reports, 2018, 38(6).
Part A, vol. 85, no. 3, pp. 214–230, 2014. [30] Bora K, Chowdhury M, Mahanta LB, et al. Automated
[11] Y.-F. Chen, P.-C. Huang, K.-C. Lin, H.-H. Lin, L.-E. Wang, C.-C. classification of Pap smear images to detect cervical dysplasia[J].
Cheng, T.-P. Chen, Y.-K. Chan, and J. Y. Chiang, “Semi-automatic Computer methods and programs in biomedicine, 2017, 138: 31-47.
segmentation and classification of pap smear cells,” IEEE Journal [31] Su J, Xu X, He Y, et al. Automatic detection of cervical cancer cells
of Biomedical and Health Informatics, vol. 18, no. 1, pp. 94–108, by a two-level cascade classification system[J]. Analytical Cellular
2014. Pathology, 2016, 2016.
[12] T. Chankong, N. Theera-Umpon, and S. Auephanwiriyakul, [32] William W, Ware A, Basaza-Ejiri AH, et al. A pap-smear analysis
“Automaticcervical cell segmentation and classification in pap tool (PAT) for detection of cervical cancer from pap-smear
smears,” Comput. Meth. Programs Biomed., vol. 113, no. 2, pp. images[J]. Biomedical engineering online, 2019, 18(1): 16.
539–556, 2014. [33] Ashok B, Aruna P. Comparison of Feature selection methods for
[13] Jantzen J, Norup J, Dounias G, et al. Pap-smear benchmark data for diagnosis of cervical cancer using SVM classifier[J]. Int. J. Eng. Res.
pattern classification[J]. Nature-inspired Smart Information Appl, 2016, 6: 94-99.
Systems (NiSIS 2005), 2005: 1-9. [34] Adem K, Kiliçarslan S, Cömert O. Classification and diagnosis of
[14] Y. Marinakis, M. Marinaki, and G. Dounias, “Particle swarm cervical cancer with stacked autoencoder and softmax
optimization for pap-smear diagnosis,” Expert Systems with classification[J]. Expert Systems with Applications, 2019, 115:
Applications, vol. 35, no. 4, pp. 1645–1656, 2008. 557-564.
[15] Marinakis Y, Dounias G, Jantzen J. Pap smear diagnosis using a [35] Abdoh S F, Rizka M A, Maghraby F A. Cervical cancer diagnosis
hybrid intelligent scheme focusing on a genetic algorithm based using random forest classifier with SMOTE and feature reduction
feature selection and nearest neighbor classification[J]. Computers techniques[J]. IEEE Access, 2018, 6: 59475-59485.
in Biology and Medicine, 2009, 39(1): 69-78. [36] Hussain E, Mahanta LB, Das C R, et al. A comprehensive study on
[16] Nanni L, Lumini A, Brahnam S. Local binary patterns variants as the multi-class cervical cancer diagnostic prediction on pap smear
texture descriptors for medical image analysis[J]. Artificial images using a fusion-based decision from ensemble deep
intelligence in medicine, 2010, 49(2): 117-125. convolutional neural network[J]. Tissue and Cell, 2020:101347.
[17] Guo Y, Zhao G, PietikäInen M. Discriminative features for texture [37] Pirovano A, Almeida L G, Ladjal S. Regression Constraint for an
description[J]. Pattern Recognition, 2012, 45(10): 3834-3843. Explainable Cervical Cancer Classifier[J]. arXiv preprint
[18] Sokouti B, Haghipour S, Tabrizi A D. A framework for diagnosing arXiv:1908.02650, 2019.
cervical cancer disease based on feedforward MLP neural network [38] Lin H, Hu Y, Chen S, et al. Fine-grained classification of cervical
and ThinPrep histopathological cell image features[J]. Neural cells using morphological and appearance-based convolutional
Computing and Applications, 2014, 24(1): 221-232. neural networks[J]. IEEE Access, 2019, 7: 71541-71549.
[19] Zhao M, Wu A, Song J, et al. Automatic screening of cervical cells
using block image processing[J]. BioMedical Engineering OnLine,
2016, 15(1): 14.

VOLUME XX, 2017 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

Authors
Hua Dun graduated from Hebei North
Suxiang Yu received the master degree University with a bachelor's degree in
in the Anatomy and Histology and clinical Medicine in June 2007. From
Embryology from Chengde Medical 2008 to now, she has been working in
College in 2009, graduated in clinical the Pathology Department of Tang
medicine from Hebei Medical County People's Hospital. From
University in June 1996. She is August 2014 to August 2015, She
currently a Deputy Director in Baoding studied for one year in the Pathology
Center for Quality Control in Department of Peking University
Pathology. She was also appointed a Third Hospital and obtained the
member of Hebei Anti-cancer certificate of completion.
Association and Hebei Maternal and
Child Health Association in 2016. Her current research
interests include early diagnosis and prevention of tumor. Now ZHANG SHUAI received the B.Eng.
she works in the Pathology Department of the Fourth Central degree in software engineering from
Hospital of Baoding City. From August 2016 to August 2017, Beijing University of Posts and
she studied for half a year in the Pathology Department of Telecommunications. He is going to
China-Japan Friendship Hospital. studying for a master's degree in
artificial intelligence in University of
Manchester during from October
Xinxing Feng received the 2020 to June 2021. His current
research interests include machine
M.D. degree from Peking learning, computer vision and
nion Medical College. He software engineering.
is currently an associate
professor in Fuwai Ruihong Zhang, chief nurse;
graduated from Shanxi Medical
Hospital, National Center University in 2009, with a master's
for Cardiovascular degree. She is currently an
President Assistant of the hospital
Diseases, Chinese and Chief of Science and Education.
Academy of Medical Sciences and Peking Union She started working in the Fourth
Central Hospital of Baoding City in
Medical College. He is interested in the field of April 2003. From February to May
combining medicine with artificial intelligence. 2017, she passed the selection
examination of the Talent
Exchange Center of the National
Health and Family Planning Commission, went to Davis
Bin Wang received the bachelor Medical Center attached to the University of California to
degree in clinical Medicine from study, participated in wound management and health
HeBei Medical Univesity in 2007 leadership training programs, and was highly praised by
year,the master degree in pathology American medical staff. She is currently the deputy chairman
and pathophysiology from ShanXi of Baoding Nursing Research Working Committee, the deputy
Medical Univestity in 2010 year.Since chairman of Baoding Youth Talent Committee, the member of
July 2010, He has been working in the Hebei Nursing Association, and the third-level talent of Hebei
Department of Pathology ,the Fourth Province.
Central Hospital of Baoding city,
Hebei province,No.28 Xiangyang
North Street,Tang County China.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3060447, IEEE Access

Xin Huang received his B. S., M.

S. and Ph.D. degrees from Harbin
Institute of Technology, Harbin,
China in 2004, 2006 and 2010,
respectively. He worked as a
Postdoctoral Fellow with Solar
Activity Prediction Center,
National Astronomical
Observatories, Chinese Academy
of Sciences (NAOC) from 2010 to
2012. He joined Solar Activity
Prediction Center, NAOC in 2012.
Now he is an associate professor.
His research interests are focused on solar activity prediction
and medical data mining.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/