Automatic Classiﬁcation of Cervical Cell Patches based on

Non-geometric Characteristics

Douglas Wender A. Isidoro

, Cl

audia M. Carneiro

, Mariana T. Resende

, F

atima N. S. Medeiros

Daniela M. Ushizima

and Andrea G. Campos Bianchi

Computer Department, Universidade Federal de Ouro Preto, Ouro Preto, Brazil

Clinical Analysis Department, Universidade Federal de Ouro Preto, Ouro Preto, Brazil

Teleinformatics Engineering Department, Universidade Federal do Cear

a, Cear

a, Brazil

Lawrence Berkeley National Laboratory, Universidade de Berkeley, Calif

ornia, U.S.A.

Keywords:

Pattern Recognition, Texture Features, Cervical Cell, Classiﬁcation, Pap Smear.

Abstract:

This work presents a proposal for an efﬁcient classiﬁcation of cervical cells based on non-geometric char-

acteristics extracted from nuclear regions of interested. This approach is based on the hypothesis that the

nuclei store much of the information about the lesions in addition to their areas being more visible even with

a high level of celular overlap, a common fact in the Pap smears images. Classiﬁcation systems were used in

two and three classes for a set of real images of the cervix from a supervised learning method. The results

demonstrate high classiﬁcation performance and high efﬁciency for applicability in realistic environments,

both computational and biological.

1 INTRODUCTION

With approximately 530,000 new cases per year

worldwide, cervical cancer is the third most frequent

tumor among the female population, behind breast

and colorectal cancer. It is responsible for 265,000

deaths per year, being the fourth most frequent cause

of cancer death in women (World Health Organiza-

tion, 2018). However, this neoplasm has a slow devel-

opment, which increases the chance of cure when the

precursor lesions are identiﬁed early by cytopatholog-

ical examination.

The most eminent screening test for the detection

of cervical cancer in its early stages is Pap smears,

introduced by George Papanicolaou. Its polarization

was due to its low cost and easy access in develop-

ing countries, as well as its great capacity for dif-

ferentiating the types of lesions found, which is ex-

tremely important for the correct diagnosis. Although

this technique is widely used, it presents signiﬁcant

false negative, false positive and unsatisfactory re-

sults, with causes attributed to different stages of the

process. Therefore, in Brazil, it is recommended that

cytopathology laboratories implement a quality mon-

itoring with the objective of improving the perfor-

mance of the results obtained in this exam.

Based on this objective, new technologies such as

cytology in liquid basis, molecular biology and the

semi-automation of the reading of the cytopathologi-

cal examination have been implemented in the health

market. The beneﬁts of liquid cytology do not jus-

tify its high cost, which makes it inaccessible to large

laboratories, including those that serve the Uniﬁed

Health System in Brazil (SUS). Efﬁcient diagnosis

from molecular biology is only really effective when

jointly used in with the cytopathological examination.

Thus, conventional cytology is the technique chosen

to perform the Pap test, not excluding the need for as-

sociation with other strategies to improve the quality

of cervical cancer screening.

The cytopathological interpretation is performed

by qualiﬁed specialists and is based on the visual

recognition, in optical microscope, of the alterations

of the nucleus, cytoplasm, and other celular informa-

tion present in the smears. These changes are still as-

sociated with clinical opinions. After the application

of this screening technique, enlarged images contain-

ing the cellular characteristics are obtained by means

of a photocopy of the Pap smears. However, due to

the high complexity of the ﬁelds analyzed microscop-

ically and consequently of the images obtained, this

process is conﬁgured as slow, exhaustive and error-

prone.

In images containing cells overlap and clusters,

Isidoro, D., Carneiro, C., Resende, M., Medeiros, F., Ushizima, D. and Bianchi, A.

Automatic Classiﬁcation of Cervical Cell Patches based on Non-geometric Characteristics.

DOI: 10.5220/0009172208450852

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 5: VISAPP, pages

845-852

ISBN: 978-989-758-402-2; ISSN: 2184-4321

845

Figure 1: Isolated (A) and overlapping cells (B) present on

Pap smears. Note that the area of the cytoplasm is clearly

visible in A while in B the exact edges of the cytoplasm

areas are ambiguous.

detection of the cytoplasmic boundary is a difﬁcult

problem as exempliﬁed in Figure 1 and to date there

are no efﬁcient methods capable of performing this

automatically. However, the detection and segmenta-

tion of nuclei in images containing certain degree of

cells overlap and clusters have been successfully ad-

dressed by several studies (Plissiti et al., 2011; Plis-

siti et al., 2011; Sobrevilla et al., 2010). The re-

sults, which mostly use images from the cytopatho-

logical examination of the liquid-based cytology, in-

dicate that the nuclear characteristics may be sufﬁ-

cient to differentiate lesions present in cervical im-

ages. Regarding classiﬁcation of lesions, some au-

thors also good results with nuclei but using isolated

cells (Lorenzo-Ginori et al., 2013).

Based on the above, it is noted that the measures

adopted to guarantee the quality of the cytological ex-

amination are still insufﬁcient. Therefore, based on

the viability provided by the use of computational

methods to improve this process, it is evident the need

to investigate technological strategies that seek to im-

prove health services in the scenario of screening for

this neoplasm. In this sense, the aim of this work is to

perform the classiﬁcation of real Pap smears images

by extracting non-geometric features present in cell

nuclei, excluding the use of segmentation techniques

or neural networks. The idea is to generate something

similar to the screening done by the pathologist on

the slides, i.e., to evaluate parts of an image and to

identify regions with lesions without the cellular indi-

vidualization. From this, we intend to investigate the

stiffness necessary for the proposed system to provide

efﬁcient aid for diagnosis in a biological and medici-

nal environment.

This paper is organized as follows, Section 2

presents literature based on cervical cell classiﬁca-

tion. Section 3 details the used methodology describ-

ing all its steps. The results are discussed in Section 4.

The conclusions are exposed in Section 5. And then,

the further works are introduced in Section 6.

2 RELATED WORKS

The literature presents some techniques for the classi-

ﬁcation of cervical cellular microscopy images. The

most widely used public image base among them is

Herlev, which contains 917 Pap Smears images and is

provided by the University Hospital of Herlev, Den-

mark (Jantzen et al., 2005). Its data set consists of 7

classes of cervical cytology images with single cell,

being: superﬁcial squamous, squamous intermediate,

columnar, moderate dysplasia, moderate dysplasia,

severe dysplasia and carcinoma in situ. Some works

that are based on this collection of images are men-

tioned below:

Mariarputham and Stephen (2015) presents a cell

classiﬁcation system, based on texture information

extracted from the nucleus and cytoplasm, using neu-

ral networks and supervised learning methods. It is

emphasized that there is no unique set of characteris-

tics capable of providing efﬁcient diagnostics for any

and all classes of a proposed classiﬁcation system. In

addition, SVM algorithm is presented as an efﬁcient

solution for classiﬁcation.

Lakshmi and Krishnaveni (2014) describes a

method for automated extraction of multiple charac-

teristics of nuclei and cytoplasms, using a paramet-

ric probability density function associated with Maxi-

mization Expectation algorithm and clustering. Their

results conﬁrm the efﬁciency of the proposed system

to differentiate lesions between low grade (LSIL) and

high grade (HSIL).

Singh et al. (2015) proposes a classiﬁcation tech-

nique using Random Forest. For this, it considers

the texture characteristics of the nucleus and cyto-

plasm, investigates the efﬁciency of these character-

istics for the recognition of normal and altered cells

and presents metrics to measure the efﬁciency of the

system.

Walker et al. (1994) despite using his own

database, presents preliminary results for the classiﬁ-

cation of cell nuclei using textural features of the Gray

Level Co-Occurrence Matrix (GLCM) and linear dis-

criminant analysis to reduce the dimensionality of the

characteristics extracted. His studies demonstrate that

the texture characteristics extracted from GLCM pro-

vide efﬁcient means for discriminating normal and al-

tered cervical cells.

All the previously mentioned works use the ex-

traction of texture characteristics associated with the

extraction of geometric characteristics and a pre-

processing that contains cellular segmentation. In a

different way, Plissiti and Nikou (2012) presents a

framework for efﬁcient classiﬁcation of cervical cells

into normal and altered categories based exclusively

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

846

on the extraction of texture characteristics from the

core area. Non-supervised classiﬁers and non-linear

clustering techniques are used, as well as genetic al-

gorithms for resource selection. Also examined is

the efﬁciency of nonlinear dimensionality reduction

schemes to produce an accurate representation of the

multiple cell characteristics. This provides results

that allow to analyze the efﬁciency of systems based

on the exclusive extraction of characteristics coming

from the nuclei when compared to the systems that

use extraction of characteristics of the complete cell.

In the state of the art, there are methods that em-

ploy Deep Learning, as in Arajo et al. (2019). This

paper presents a computational tool for cytological

analysis using Deep Learning (CNN) techniques for

cell segmentation in conventional Pap smears im-

ages containing a high degree of cell overlap. Its

results demonstrate high efﬁciency for the proposed

approach, as well as robustness for the presence and

interference of neutrophils, noises and other artifacts

very common in Pap smears that compromise auto-

mated classiﬁcation. Despite the effectiveness, tech-

niques that use Deep Learning and neural networks,

present some dysfunctions: they need a large num-

ber of images for training; present high computational

cost; and, mainly, do not allow the identiﬁcation of the

attributes used and their respective amounts. Some-

thing that makes it impossible to correlate the com-

putational technique with the biological context and

thus leads us to seek alternative solutions, as well as

the system presented here.

Differently from the techniques mentioned above,

our proposal uses a database of real cervical cytology

images with multiple cells obtained from pap smears

images. In addition, the segmentation and extraction

of geometric characteristics are not used, in order to

allow the classiﬁcation of images where it is not pos-

sible to identify cytoplasmic and nuclear boundaries.

3 METHODOLOGY

This section describes the stages of development of

this article. Section 3.1 speciﬁes the image database

used. Section 3.2 describes the pre-processing step to

which the images are submitted. Section 3.3 presents

the textural features extracted from the previously se-

lected set of images. Section 3.4 introduces the super-

vised learning method used for classiﬁcation, as well

as describing the system of class division chosen for

this system. Finally, Section 3.5 describes the process

of analyzing the results obtained on the basis of sta-

tistical metrics. The proposed system architecture is

represented in Figure 2.

Figure 2: Architecture of the proposed system.

3.1 Database

All images used were obtained from the the Brazil-

ian Health System (Arajo et al., 2019). This database

consists of 11473 cellular nuclei obtained from 400

1392x1040 images. All of them are scanned from the

use of an optical microscope associated with a 40x

magniﬁcation camera and individually represent seg-

ments of the slides that may have different numbers

and cell types. This private database is a product of a

multidisciplinary team of researchers in computer sci-

ence and biology who provides a collection of real im-

ages of human cervical cytology smears representing

a range of different cervical cell lesions interpreted by

independent cytopathologists, which are listed below:

1. Normal cells;

2. Atypical squamous cells of undetermined signiﬁ-

cance (ASC-US);

3. Atypical squamous cells, cannot exclude a high-

grade lesion (ASC-H);

4. Low-grade squamous intraepithelial lesion

(LSIL);

5. High-grade squamous intraepithelial lesion

(HSIL);

6. Scamous cell carcinoma (SCC);

3.2 Pre-processing and Training

Database

As mentioned, since we do not use nuclei or cyto-

plasm segmentation, all parts of the image are used

for the training and construction of the classiﬁcation

model. In this preprocessing step the database images

are cut out in 50x50 pixels throughout their size, as

exempliﬁed in Figure 3.

These cuts are made in an orderly manner from

left to right and top to bottom. However, it is arbi-

trary in relation to the exact position of the cells and

Automatic Classiﬁcation of Cervical Cell Patches based on Non-geometric Characteristics

847

Figure 3: Example of preprocessing step. The numbering

represented in this image is unrealistic only for demonstra-

tion purposes, as well as the size of the cutouts.

nuclei, as illustrated in Figure 3. The deﬁnition of

the information for the training is very important be-

cause they accurately compose the set of information

regarding each lesion and lead to a good accuracy in

the result. Accordingly, the pre-processing step re-

sults in reduced-size images which:

• They comprise sufﬁcient nuclear information to

be encompassed by any of the classes of lesions

described below;

• They comprise insufﬁcient nuclear fragments to

be considered within some class of lesions and are

thus categorized as noises;

• They comprise cell overlaps and are only enclosed

within injury classes if they include at least 60%

of the nuclear information of any of the cells in-

volved;

• They do not comprise any nuclear information

and therefore are also included in the noise cat-

egory;

To verify if there is enough information in a clip-

ping to match one of the nuclear classes of the clas-

siﬁcation system, the data provided by the database

is used in the training step regarding the location of

the nuclei within the images. Based on the positions

of the nuclei, a comparison is made between cut-outs

made from the central positions of all nuclei present in

the image and the arbitrary cut-off performed by the

pre-processing at that time, as illustrated by Figure 4.

In addition, this cut-out around the central position

(x, y) of an image core is used as input to the train-

ing and testing step, along with the arbitrary cut-offs

obtained, in order to allow the classiﬁer to also learn

how to categorize cores as a whole.

As a consequence, the overlap value of these im-

ages is obtained, which, if having a value equal to

zero, is indicated as position cutouts that do not cor-

respond to image cores. If a value greater than zero is

detected, the juxtaposition of the cutouts is detected

and in the speciﬁc case of having a value greater than

or equal to 60% of the area of the cut, it is determined

Figure 4: Example of a comparison between cuts, where:

(A) represents the cut form a core central position; (B) rep-

resents the cut being performed by a preprocessing at that

instant (C) represents the intersection between two cuts,

which, in this case, have higher or equal value to 60% of

their individual areas.

that it has enough information to be framed as a cell

nucleus region and is identiﬁed as one of the classes

of lesions, different from the noises.

The proposed trimming creates a new unit delim-

ited in the region near the cellular nucleus, with less

information, alternatively for the segmentation that is

not used. The size of this cut was empirically de-

ﬁned from a series of tests and parameter adjustments

where the best results were presented by the images in

the dimension chosen here. All cutouts obtained from

pre-processing are used as input to the sorting step. It

should be noted that the classiﬁcation step does not

use this information regarding the position of the nu-

cleus, but rather a comparison between the attributes

of a region of the image with those used in the train-

ing.

3.3 Feature Extraction

From the study presented in Section 2, a subset of

characteristics with a high discriminative capacity to

be extracted from the set of cutouts deﬁned as training

was deﬁned. All selected features refer to information

related to the texture of the images and do not undergo

segmentation preprocessing to be obtained due to the

research objective. Are they:

• Average Intensity;

• Maximum Intensity Value;

• Minimum Intensity Value;

• Local Binary Patern (LBP);

• Histogram of Oriented Gradients (HOG);

• Gray-Level Co-Occurrence Matrix (GLCM);

• 7 Hu Moments;

• Haralick’s Texture Features (Haralick, 1979):

– Angular Second Moment (Energy);

– Contrast;

– Correlation;

– Variance;

– Entropy;

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

848

– Maximal Correlation Coefﬁcient;

– Inverse Difference Moment (Homogeneity);

– Sum Average, Sum Variance and Sum Entropy;

– Difference Variance and Difference Entropy;

– Information Measure of Correlation I and II.

3.4 Classiﬁcation and Testing

The manual classiﬁcation of the entire dataset of the

base used was done by specialists of the area and was

used as terrestrial truth. For the classiﬁcation pro-

posed here, two and three classes systems were used,

shown in Figure 5 and described as follows:

• 2Class: In this system, the lesions are grouped

into two classes. In the ﬁrst class only normal

cells and in the second lesions ASC-US, ASC-H,

LSIL, HSIL and SCC, forming the class of altered

cells. The classiﬁcation using this system aims to

classify images only from the presence or absence

of lesion.

• 3Class: In this system, the lesions are grouped

into three classes. In the ﬁrst class, there are only

normal cells. In the second, there are the lesions

of light severity ASC-US and LSIL. And ﬁnally,

in the third class, there are the lesions of high

ASC-H, HSIL and SCC. Throughout the remain-

der of this document, the second class is referred

to as mild injuries and the third class is referred

to as serious injuries. This proposal aims to clas-

sify images based on the diagnosis and treatment

given to the patients detected with these groups of

lesions. Where: patients without lesion (normal

class) after receiving two negative results of the

examination, only need to repeat it after 3 years;

Patients with light-class lesions need follow-up

and should repeat the exam within 6 months or

1 year; patients with severe class injuries require

special care and must perform colposcopy and/or

biopsy.

Figure 5: Classiﬁcation system used.

In both systems a further class is included that

includes cutouts that do not have relevant classiﬁca-

tion information, such as blade bottoms or other non-

nuclear elements. The purpose of this class is to allow

the classiﬁer to learn to ignore information that is not

important for cell sorting. This class will be referred

to as noises in the remainder of this document.

For the classiﬁcation was used the algorithm Sup-

port Vector Machine (Bishop, 2006). This method of

supervised learning is a formally discriminative clas-

siﬁer that identiﬁes and constructs a model with a

class separation hyper plane. It receives as input a set

of data and predicts, for each given input, which of the

possible classes the input is part of, making it a non-

probabilistic linear classiﬁer. In the training stage, all

the image cutouts deﬁned in Section 3.2 were used. In

addition, K-Fold cross validation is applied to ensure

the generality and impartiality of the results obtained,

using K=10%.

3.5 Evaluation

For the evaluation of the results and classiﬁcation per-

formance, some metrics were extracted. First, for the

calculation of statistical measures, the number of true

positives (TP), which refer to altered nuclei correctly

classiﬁed as altered, true negatives (TN), which re-

fer to normal nuclei correctly classiﬁed as normal,

false positives (FP), which refer to normal nuclei erro-

neously classiﬁed as altered and false negatives (FN),

which refer to altered nuclei erroneously classiﬁed as

normal. The statistical measures are described below:

3.5.1 Precision

Is the percentage of data that is correctly classiﬁed

for the class to which it truly belongs, i.e. normal as

a normal class or altered as altered class in the 2-class

problem. In the 3-class problem, it is applied directly

to the correctly classiﬁed data in the normal classes,

mild lesions and severe lesions. Precision is deﬁned

as:

Precision =

T P + T N

T P + FN + T N + FP

(1)

3.5.2 Sensitivity

Is the percentage of changed data that is correctly

classiﬁed as changed (true positive). The sensitivity

is deﬁned as:

Sensitivity =

T P

T P + FN

(2)

The sensitivity calculation in the 2-class problem

is simple. However, in the 3-class problem, we need

to regroup ASC-US, LSIL, ASC-H, HSIL and SCC to

form a single class of altered cells. After that, the sen-

sitivity calculation is the same as in the 2-class prob-

lem.

Automatic Classiﬁcation of Cervical Cell Patches based on Non-geometric Characteristics

849

3.5.3 Speciﬁcity

Is the percentage of normal data that is correctly clas-

siﬁed as normal (true negative). Speciﬁcity is deﬁned

as:

Speci f icity =

T N

T N + FP

(3)

Similar to the Sensitivity calculation, the compu-

tation of Speciﬁ- city in the 2-class problem is sim-

ple, whereas ASC-US, LSIL, ASC-H, HSIL and SCC

need to be regrouped as a class of altered cells in the

3-class problem.

3.5.4 F1 Score

Is a relationship between precision and sensitivity to

provide a single, balanced measurement for the sys-

tem. The F1 Score is deﬁned as:

F1 = 2 ∗

Precision ∗ Sensitivity

Precision + Sensitivity

(4)

4 RESULTS AND DISCUSSION

The classiﬁcation was made from texture character-

istics extracted from images obtained from the Pap

smears examination. Different combinations of char-

acteristics were experienced from the use of Principal

Component Analysis (PCA) as a dimensionality re-

duction technique. The best results were presented

using all the characteristics described in Section 3.3.

The step-by-step procedure of the system pro-

posed here is shown in Figure 2. As discussed in Sec-

tion 3.4, two distinct classiﬁcation experiments are

proposed, one with two classes (2Class), investigat-

ing the presence or not of image damage, and another

with three classes (3Class), allowing differentiation

not only between the presence or absence of lesions,

but also between mild or severe lesions.

The performance of several classiﬁcation algo-

rithms was analyzed and compared. The most ef-

ﬁcient method for separating the data set was the

SVM with accuracy of 89.7% for the classiﬁcation in

two classes and 85.1% for the classiﬁcation in three

classes. Other results obtained are presented in Table

1 and Table 2. It is important to note that, in order

to calculate these metrics, the information about the

noise class was disregarded, that is, any values ob-

tained from the confusion matrix involving the noise

class were not accounted for.

In order to provide effective aid in the detection

of lesions in images obtained from Pap smears, the

effectiveness of the results obtained was investigated

Table 1: Results for 2 Class patches classiﬁcation.

Metrics Values

Precision 0.897

Recall 0.917

Speciﬁcity 0.883

F1 Score 0.907

Table 2: Results for 3 Class patches classiﬁcation.

Metrics Values

Precision 0.851

Recall 0.954

Speciﬁcity 0.778

F1 Score 0.899

in both computational and biological bias. According

to the experts who performed the manual classiﬁca-

tion of the database, as the objective of the system is

to provide a classiﬁcation tool or preliminary rank-

ing of the images obtained in the slides, not exclud-

ing the function of the cytopathologist at the end of

the structure to make the diagnosis ofﬁcial in the sce-

nario of a possible presence of injury, but rather pro-

viding support for decision-making, emphasizing that

the main problem that must be tackled within the pro-

posed structure are the false negative classiﬁcations,

i.e. do not ﬁnd lesions that actually exist.

In these situations, the diagnosis of an injury is

not detected, causing the patient to create a false tran-

quility and remain at risk of developing the cancer.

The experts also explain that false positive diagnoses

are not essential problems in the context of the real

application of this system, because these images will

be analyzed by the cytopathologist for ﬁnal diagno-

sis, and if the blade indicated with lesions does not

actually have them, the diagnosis will be given as nor-

mal and the patient will receive the correct diagnosis.

However, in false negative diagnoses, the blade that

contains lesions and is classiﬁed as normal is not des-

ignated for evaluation in the second instance by the

cytopathologist and in this way the system could lead

to incorrect diagnoses.

From this, several experiments were performed

and the performance of the classiﬁcation technique

was measured by varying the deﬁned parameters in

order to obtain the greatest possible decrease of the

amount of false negatives reached by the system. The

obtained results are satisfactory for the context of the

application in real environment, being: 1.87% of false

negatives in the system of 3 classes and 3.41% in the

system of 2 classes. It is also worth mentioning that,

throughout the development of this article, the used

techniques were chosen in order to reduce as much as

possible the computational cost of the proposed struc-

ture. This is due to the fact that, from a single Papan-

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

850

icolaou slide, are obtained between 10 to 15 thousand

images with a 40x magniﬁcation and, as in a realis-

tic environment there is too many of these slides for

analysis, in order that the structure here proposed has

functional applicability in this environment, the clas-

siﬁcation of an image must happen quickly. Making

elementary techniques stand out in relation to robust

techniques.

5 CONCLUSION

Precisely classifying cell nuclei in real images of Pap

smear is a necessary condition to provide more ac-

curate and reliable diagnostics. Since in this type

of image the information about nuclear and cytoplas-

mic boundaries is complex and computationally in-

feasible, this paper presents a structure for the clas-

siﬁcation of real cervical cytology images from the

extraction of non-geometric textural features without

the use of learning from neural networks or from seg-

mentation algorithms.

The proposed structure has realistic application to

help in the diagnosis and screening of cervical cancer.

A supervised learning classiﬁer, SVM, was used for

classiﬁcation and system performance was measured

with actual cervical cytology images, created from the

Pap smear, provided by a private searchable imaging

database.

The results obtained by our experiments indicate

the optimum performance of the proposed system in

the process of categorization of lesions present in Pap

smears images, presenting high precision and low

false negative index. This implies that the efﬁcient

characterization of an actual cervical cytology image

is feasible only with the use of non-geometric extrac-

tion features.

In the medical and biological context, the pro-

posed structure serves as a preliminary Pap smears

classiﬁer that determines the likelihood of lesions in

an image and signals them to be analyzed by a cy-

topathologist when necessary and provide a deﬁnitive

diagnosis of more practical, efﬁcient and with quality

assurance. In addition, correlating biology and com-

putation, pertinent conclusions can be inferred regard-

ing the rigidity or ﬂexibility required for a system to

have effective applicability in real environments.

6 FUTURE WORKS

As future works we intend to investigate the efﬁciency

of the proposed system when applied in a larger num-

ber of real images not cataloged and to construct a

system that allows the return of these regions to the

analysis of the cytopathologist. Another important

factor to be examined are the sets of descriptors, per-

forming a more systematic investigation for the in-

clusion of information that allows the increase of the

precision of the method.

ACKNOWLEDGEMENTS

The authors thanks Conselho Nacional de De-

senvolvimento Cient

ıﬁco e Tecnol

ogico (PIBITI-

CNPq), Universidade Federal de Ouro Preto (UFOP),

Fundac¸

ao de Amparo

a Pesquisa do Estado de Mi-

nas Gerais (PPSUS-FAPEMIG/APQ-03740-17), the

Moore-Sloan Foundation, and Ofﬁce of Science,

of the U.S. Department of Energy under Contract

No. DE-AC02-05CH11231 for also supporting this

research. Any opinion, ﬁndings, and conclusions or

recommendations expressed in this material are those

of the authors and do not necessarily reﬂect the views

of the Department of Energy or the University of Cal-

ifornia.

REFERENCES

Arajo, F. H., Silva, R. R., Ushizima, D. M., Rezende, M. T.,

Carneiro, C. M., Bianchi, A. G. C., and Medeiros,

F. N. (2019). Deep learning for cell image segmen-

tation and ranking. Computerized Medical Imaging

and Graphics, 72:13 – 21.

Bishop, C. M. (2006). Pattern Recognition and Ma-

chine Learning (Information Science and Statistics).

Springer-Verlag, Berlin, Heidelberg.

Haralick, R. M. (1979). Statistical and structural approaches

to texture. Proceedings of the IEEE, 67(5):786–804.

Jantzen, J., Norup, J., Dounias, G., and Bjerregaard, B.

(2005). Pap-smear benchmark data for pattern clas-

siﬁcation. Nature Inspired Smart Information Systems

(NiSIS).

Lakshmi, G. K. and Krishnaveni, K. (2014). Multiple fea-

ture extraction from cervical cytology images by gaus-

sian mixture model. In 2014 World Congress on Com-

puting and Communication Technologies, pages 309–

311.

Lorenzo-Ginori, J. V., Curbelo-Jardines, W., L

opez-

Cabrera, J. D., and Huergo-Su

arez, S. B. (2013). Cer-

vical cell classiﬁcation using features related to mor-

phometry and texture of nuclei. In Ruiz-Shulcloper,

J. and Sanniti di Baja, G., editors, Progress in Pat-

tern Recognition, Image Analysis, Computer Vision,

and Applications, pages 222–229, Berlin, Heidelberg.

Springer Berlin Heidelberg.

Mariarputham, E. J. and Stephen, A. (2015). Nominated

texture based cervical cancer classiﬁcation. In Com-

putational and Mathematical Methods in Medicine.

Automatic Classiﬁcation of Cervical Cell Patches based on Non-geometric Characteristics

851

Plissiti, M. E. and Nikou, C. (2012). Cervical cell clas-

siﬁcation based exclusively on nucleus features. In

Campilho, A. and Kamel, M., editors, Image Analysis

and Recognition, pages 483–490, Berlin, Heidelberg.

Springer Berlin Heidelberg.

Plissiti, M. E., Nikou, C., and Charchanti, A. (2011). Au-

tomated detection of cell nuclei in pap smear im-

ages using morphological reconstruction and cluster-

ing. IEEE Transactions on Information Technology in

Biomedicine, 15(2):233–241.

Plissiti, M. E., Nikou, C., and Charchanti, A. (2011). Com-

bining shape, texture and intensity features for cell nu-

clei extraction in pap smear images. Pattern Recogni-

tion Letters, 32(6):838 – 853.

Singh, D., Verma, A., Aneja, M., and Singh, B. (2015). Cer-

vical cell classiﬁcation using random forests. Interna-

tional Conference on Medical Physics and Biomedical

Engineering.

Sobrevilla, P., Montseny, E., Vaschetto, F., and Lerma, E.

(2010). Fuzzy-based analysis of microscopic color

cervical pap smear images: Nuclei detection. Interna-

tional Journal of Computational Intelligence and Ap-

plications, 9:187–206.

Walker, R. F., Jackway, P., Lovell, B., and Longstaff, I. D.

(1994). Classiﬁcation of cervical cell nuclei using

morphological segmentation and textural feature ex-

traction. In Proceedings of ANZIIS ’94 - Australian

New Zealnd Intelligent Information Systems Confer-

ence, pages 297–301.

World Health Organization (2018). Global cancer statistics.

http://gco.iarc.fr/. Last checked on Oct 23, 2018.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

852