AnAutomaticMassScreeningSystemforCervicalCancer ...single normal cell and 160 image patches from each ab-normal cell. is transformation yields relative normal distribution, as the

Research ArticleAn Automatic Mass Screening System for Cervical CancerDetection Based on Convolutional Neural Network

Aziz-ur -Rehman 1 Nabeel Ali 2 ImtiazA Taj 2 Muhammad Sajid 3

and Khasan S Karimov14

1Department of Electrical Engineering GIK Institute of Engineering Sciences and Technology Topi 23640 District SwabiKhyber Pakhtunkhwa Pakistan2Department of Electrical Engineering Capital University of Science and TechnologyIslamabad Expressway Kahuta Road Zone-V Islamabad Pakistan3Department of Electrical Engineering Mirpur University of Science and Technology (MUST) Mirpur 10250 Pakistan4Centre for Innovative and New Technologies of Academy of Sciences of the Republic of Tajikistan Rudaki Ave 33Dushanbe 734015 Tajikistan

Correspondence should be addressed to Aziz-ur -Rehman azizur80hotmailcom

Received 19 May 2020 Revised 30 September 2020 Accepted 8 October 2020 Published 28 October 2020

Academic Editor Jose Merodio

Copyright copy 2020 Aziz-ur -Rehman et al )is is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

Cervical cancer is the fourth most common type of cancer and is also a leading cause of mortality among women across the worldVarious types of screening tests are used for its diagnosis but the most popular one is the Papanicolaou smear test in which cellcytology is carried out It is a reliable tool for early identification of cervical cancer but there is always a chance of misdiagnosisbecause of possible errors in human observations In this paper an auto-assisted cervical cancer screening system is proposed thatuses a convolutional neural network trained on Cervical Cells database )e training of the network is accomplished throughtransfer learning whereby initializing weights are obtained from the training on ImageNet dataset After fine-tuning the networkon the Cervical Cells database the feature vector is extracted from the last fully connected layer of convolutional neural networkFor final classificationscreening of the cell samples three different classifiers are proposed including Softmax regression (SR)Support vector machine (SVM) and GentleBoost ensemble of decision trees (GEDT))e performance of the proposed screeningsystem is evaluated for two different testing protocols namely 2-class problem and 7-class problem on the Herlev databaseClassification accuracies of SR SVM and GEDT for the 2-class problem are found to be 988 995 and 996 respectivelywhile for the 7-class problem they are 9721 9812 and 9885 respectively )ese results show that the proposed systemprovides better performance than its previous counterparts under various testing conditions

1 Introduction

Cervical cancer is the leading cause of cancer-related deathsin females It arises from the cervix ie the lower andnarrow end of the uterus as shown in Figure 1 It starts dueto abnormal growth of cells that have the ability to spreadinto other parts of the body Human Papilloma Virus (HPV)infection is the major risk factor for cervical cancer )ereare no symptoms in the beginning of the disease while withthe passage of time symptomsmay include abnormal vaginalbleeding pelvic pain and pain during sexual intercourse It

can be diagnosed earlier through regular medical check-ups[1]

)ere are many diagnostic tests for cervical canceridentification Papanicolaou (PAP) smear is the mostcommonly used test for cervical cancer screening worldwideIn the conventional PAP smear procedure [2] a speculum isinserted into vagina to widen the walls so that vaginal smearcan be viewed Several weeks are required to prepare for thefinal results of the PAP smear test )e process is time-consuming and laborious It requires microscopic exami-nation of hundreds of thousands of cells for the diagnosis of

HindawiMathematical Problems in EngineeringVolume 2020 Article ID 4864835 14 pageshttpsdoiorg10115520204864835

precancerous and cancerous cells In every 10 to 15 positivecases there is a chance of one case to be missed in con-ventional screening [3]

)e rate of the incidence of cervical cancer is lesser in theUSA and other parts of developed countries because of earlydetection and better screening methods [4] Its rate of oc-currence has been dropped by 80 since the screeningsystems are introduced in some Nordic countries In Sweden[5] it is dropped by 65 during the last four decades and theoccurrence of cervical cancer and mortality figures are stableover the last decade However improved screening systemsare still unavailable in underdeveloped countries partly dueto the complexity and tedious nature of manual screening ofabnormal cells from a cervical cytology specimen [6 7]While auto-assisted mass screening techniques can boostefficiency they are not accurate enough to be used as aprimary tool for cervical screening [8]

During the past few years extensive research has beencarried out for the development of computer-assisted au-tomated reading systems based on cell image analysis[7 9 10])emanual screening process is normally initiatedwith the collection of cervical cell samples from the uterinecervix and their placement on a glass slide After visualinspection under a microscope these are classified intodifferent categories)e shape size texture ratio of nucleusand cytoplasm are the main characteristics for the classifi-cation task Hence for an automated system the first stepmay include segmentation of images of cell samples toextract regions of interest containing single cells with nu-cleus and cytoplasm from the noncell regions )is initialsegmentation is then followed by separation of main cellcomponents including the nucleus and cytoplasm and ex-traction of their shapetextural features However theseparation of main components and shape feature extractionis not an integral part of an automated screening system asproposed schemes in the literature include both options iewith geometrical feature extraction and without priorextraction

For a system that includes prior feature extraction ac-curate segmentation of nucleus from cytoplasm in cervicalcell images is a difficult task and is prone to error thus

limiting the success of overall system )e presence of largeirregular shapes appearance dissimilarities and cell clustersbetween malignant and benign cell nucleus is the majorproblem in accurately segmenting the cytoplasm and nu-cleus Various segmentation algorithms have been proposedby researchers to segment out cell components An iterativealgorithm for assigning pixels based on a statistical criterionfunction was proposed in [11] to separate the nucleus cy-toplasm and background In another study [12 13] Gaborfilters were applied for exploiting textural variation of thecervical cells to segment out regions of interest FuzzyC-means clustering was used in [14 15] to segment thesingle cell images into nucleus cytoplasm and backgroundHowever if the overlapping cells are taken into account theclassification accuracy is decreased significantly)erefore amajority of the presented segmentation approaches [16ndash20][11 12 14] are effective in terms of their performance forsingle and clear cervical cell images only but in the case ofoverlapping cells or other shape changes they lack theperformance accuracy

To overcome this dependency on segmentation manytechniques have been proposed during the past few yearswhich do not include prior segmentation and directlyclassify the unsegmented cell images A pixel-level classifi-cation method is proposed in [21] to classify normal andabnormal cells without prior segmentation using block-wisefeature selection and extraction techniques However thevalidation accuracy of the proposed algorithm is not up tothe mark In [22] block image processing was proposed thatincludes cropping arbitrary image blocks prior to featureextraction and the cropped blocks are then classified usingSVM However in their approach arbitrary cropping couldpotentially separate a full cell into distinct patches

Recently feature representation in image classificationproblems based on deep learning methods has become morepopular [23] In particular convolutional neural networks(ConvNets) [24] have achieved unprecedented results in the2012 ImageNet Large Scale Visual Recognition Challengewhich consisted of classifying natural images in theImageNet dataset into 1000 fine-grained categories [25]Besides they have drastically increased the accuracy in thefield of medical imaging [26 27] specifically classification oflung diseases and lymph nodes in CT images [28 29] anddetecting cervical intraepithelial neoplasia based on cervi-gram images [30] or multimodal data [31] ConvNets havealso shown superior performance in the classification of cellimages for diagnosis of pleural cancer [32]

However large datasets are essential to achieve highperformance and to overcome the problem of overfittingwith ConvNets [33] )is is a major limitation in applyingConvNets to the cervical cell classification problem as in thecase of cervical cells and a limited number of annotateddatasets are available For instance the Herlev dataset [34]only contains 917 cervical cells with 675 abnormal and 242normal cells that are insufficient for ConvNets To overcomethis limitation recently image data augmentation tech-niques have been proposed to virtually increase the size oftraining datasets and reduce the problem of overfitting [25]Data augmentation can be achieved by linear transformation

Uterus Fallopian tube

Uterine lining

Vagina

CervixOvary

Figure 1 Female reproductive system Initially the cervical cancerstarts from the cervix the lower end of uterus

2 Mathematical Problems in Engineering

of the data such as mirroring scaling translations rotationand color shifting unless the information of the object in theimage is intact Transfer learning [21 22 35ndash39] is anothersolution to overcome data overfitting In transfer learning aConvNet is first trained on large-scale natural image datasetsand then can be fine-tuned to the desired dataset which islimited in the size

In this paper an automatic screening system is proposedto classify malignant and benign cell images without priorsegmentation using ConvNets Due to limited size of Herlevdatasets transfer learning is used to initialize the weights andthen fine tune on the dataset )e feature vector at a fullyconnected layer is extracted after fine-tuning and passed tovarious classifiers To show the efficacy of the proposedapproach its performance is evaluated on the Herlev datasetfor 2-class and 7-class problems Malignant and benign cellsare considered in the 2-class problem while in the 7-classproblem all seven categories of the cervical cells have beenexplored In short the research contributions of the pre-sented work are summarized as follows

(1) Our work is aimed at developing tool for automaticclassification of cervical cells using the convolutionalneural network Unlike previous methods it doesnot require prior segmentation and hand-craftedfeatures )is method automatically extracts hier-archical features embedded in the cell image for theclassification task

(2) A data augmentation technique has been consideredto avoid overfitting )e rate of overfitting has beenreduced as the data augmentation strategy is appliedto train our network)is approach is fruitful for ournetwork to learn the most discriminative features ofcervical cells and thus achieve superior classificationresults

(3) Transfer learning is also explored for pretrainingand initial weights are reassigned to another networkfor fine-tuning on cervical cell images Training fromscratch requires a large amount of labeled data whichis extremely difficult in medical diagnosis Moreoverthe designing and adjustment of the hyper-parameters are the challenging tasks with referenceto overfitting and other issues Transfer learning isthe easiest way to overcome such problems

(4) We also conduct extensive malignant and benign cellassessment experiment on the Herlev dataset Ourresults clearly demonstrate the effectiveness of theproposed convolutional neural architecture )eexperimental results are compared with recentlyproposed methods and our approach provides su-perior performance as compared with existing sys-tems for cervical cells classification

)e paper is organized as follows the proposed meth-odology is presented in Section 2 experiments and resultsare given in Section 3 result-related discussion is presentedin Section 4 and conclusion and future work are summa-rized in Section 5

2 Proposed Methodology

)e proposed automatic mass screening system for cervicalcancer detection using ConvNets is shown in Figure 2)ereare four steps (1) data collection (2) preprocessing (3)feature learning and (4) classification of cervical cells )esesteps are explained in the following sections

21DataCollection )epublicly available Herlev Pap smeardataset is used for the training and testing purpose Itcontains 917 single cervical cell images with ground truthclassification and segmentation )e cells are categorizedinto seven different classes)ese seven classes are diagnosedby doctors and cytologists to increase the reliability of thediagnosis Furthermore these seven classes are broadlycategorized into two groups ie malignant and benign )efirst class to third class is normal or benign while fourth toseventh class is abnormal or malignant )e classrsquos distri-bution is shown in Table 1

Normal and abnormal cell images are shown in Figure 3It can be seen that the size of the nucleus in malignant orabnormal cells is larger than that of the normal cells )edifficult task from classification perspective is that thenormal columnar cells have nucleus size quite similar to thatof severe nucleus and also chromatin distribution is same

22 Preprocessing Herlev dataset consists of images thatcontain multiple cells in a single image )e data pre-processing phase includes image patch extraction from theoriginal cervical cell images and augmentation of data fortraining ConvNet 2

221 Image Patch Extraction )e proposed approach likeprevious patch-based classification methods does not di-rectly operate on original images present in the Herlevdataset that contains multiple cells at a time [40ndash43] Imagepatches each containing single cell are first extracted Inorder to extract the individual cell presegmentation ofcytoplasm is required [44] )e nuclei are first detected andthen image patches of size M times M and each centered on anucleus is generated that embed not only the size and scaleinformation of the nucleus but also the textural informationof the cytoplasm surrounding the nucleus Scale and size ofthe nucleus is a very important discriminative feature be-tween malignant and benign cells

222 Data Augmentation An image data augmentationtechnique is used to virtually increase the size of trainingdataset and reduces overfitting [25] As the cervical cells areinvariant to rotations they can be rotated from 0 minus 360degree with a step angle θ In the data augmentation processNr 10 rotations with θ 36 degree translations in thehorizontal direction Nth 15 translations up to 15 pixelsfor each normal cells while in vertical direction Ntv 8translations up to 15 pixels for each abnormal cells areperformed Hence we generate 300 image patches from a

Mathematical Problems in Engineering 3

single normal cell and 160 image patches from each ab-normal cell )is transformation yields relative normaldistribution as the numbers of samples of abnormal cellimages are as large as compared to that of normal cellimages )e size of the generated image patch is set to m

128 pixels to cover the cytoplasm region )ese patches arethen upsampled to a size 256 times 256 times 3 using bilinear in-terpolation )ese upsampled image patches as shown inFigure 4 are used in ConvNet 2 for initiating layer transferand training [28]

)e malignant cells in the Herlev dataset are 3 timesmore than the benign cells )erefore it is natural that theclassifier tends to be more biased towards the majorityclass ie the malignant cells )e unfair distribution ofdata is commonly solved by normalization of data prior toclassification whereby the ratio of positive and negativesamples of data is evenly distributed [45] )is normali-zation process improves not only the convergence rate oftraining of the ConvNets but also the classification accu-racy [25] In the proposed approach the training dataset ismade balanced by unequal augmentation of benign andmalignant cells in which a higher proportion of benigntraining samples are generated as compared to malignanttraining samples

23 Feature Learning )e ConvNets can learn to dis-criminate features automatically for an underlying task Inthis work a typical deep model is used consisting of 2ConvNets named ConvNet 1 and ConvNet 2 At first thebase network ConvNet 1 is pretrained on ImageNet databasethat consists of over 15 million labeled high-resolutionimages belonging to roughly 22000 categories [46] )eimages were collected from the web and labeled throughhuman judgment using Amazonrsquos Mechanical Turkcrowdsourcing tool In all there are roughly 12 milliontraining images 50000 validation images and 150000testing images [25] ConvNet 1 contains five convolutional(conv) layers denoted as conv1ndashconv5 followed by threepooling (pool) layers denoted as (pool1) (pool2) and(pool5) and there are three fully connected (fc) layers as(fc6) (fc7) and (fc) All these layers are transferred toConvNet 2 which is the network used for feature extractionsetting its initial parameters )is new network is then fine-tuned on the single cervical cell images of the Herlev da-tabase )is procedure is shown in Figure 5

As described earlier (conv) and (pool) layers aretransferred from ConvNet 1 at the same locations toConvNet 2 Both ConvNet 1 and ConvNet 2 share the samestructure from (conv1) to (pool5) layers However the fully

ImageNetdatabase

Herlevdatabase

Trai

ning

dat

aTe

sting

dat

a

Testimage

Pre-processeddata

Pre-processing

Classifier(SR SVM GEDT)

Deep features

Aggregatedscore

ConvNet1 (pretrained)

Transfer learning

Conv layersPooling layersReLU layers

Fully connectedlayers

ConvNet2 (fine-tuned)Conv layers

Pooling layersReLU layers


Figure 2 )e process of cervical cancer detection system using ConvNets through transfer learning

Table 1 Key statistics of Herlev dataset [36]

Type Class Cell description Numbers (cells)Normal 1 Superficial squamous epithelial 74Normal 2 Intermediate squamous epithelial 70Normal 3 Columnar epithelial 98Abnormal 4 Mild squamous nonkeratinizing dysplasia 182Abnormal 5 Moderate squamous nonkeratinizing dysplasia 146Abnormal 6 Serve squamous nonkeratinizing dysplasia 197Abnormal 7 Squamous cell carcinoma in situ intermediate 150


connected layers are modified in ConvNet 2 because thenumber of output classes is different as compared toConvNet 1 Numbers of neurons in ConvNet 2 are 4096 minus

4096 minus 7 and 4096 minus 4096 minus 2 in the case of 7-class problemand 2-class problem respectively Fully connected layers ofConvNet 2 are initialized with values from randomGaussiandistributions Local response normalization for (conv1) and(conv2) is set according to the parameters in [25] Hiddenlayers are used with rectified linear unitsrsquo activation func-tion )ere are three fully connected layers in the proposednetwork ie (fc6) (fc7) and (fc) )e feature vector for thefinal classification task is selected from the (fc7) layer whichis the last layer before the output layer )e main reason toselect feature vector from (fc7) is that it contains morespecific and abstract details of the images )e dimension offeature vector extracted from (fc7) is the number of trainingsamples Xt times 4096 )e configuration of ConvNet 2 is listedin Table 2

24 Classification Deep features are extracted from theouter layer of ConvNet 2 for cervical cells classification task

)e classification score is then calculated using three dif-ferent classifiers including SR SVM and GEDT )e detailsof three classifiers are also presented as follows

241 Softmax Regression (SR) For themulticlass dataset SRis used for classification of unknown samples that are firstpreprocessed according to the described approach UnlikeConvNets SR uses cross entropy function for the classifi-cation )e sigmoid function is replaced by softmax func-tion Mathematically it is represented by the followingequation

P y jz(i)

1113872 1113873 ϕsoftmax z(i)

1113872 1113873 e

z(i)

1113936kj0 e

z(i)

k

(1)

where we define the network input z as

z w0x0 + w1x1 + middot middot middot + wnxn 1113944n

i0wixi w

Tx (2)

where w is a weight vector x is the feature vector of trainingsample and w0 is the bias )is softmax function computesthe probability score that training sample xi belongs to classj given the network z )e probability score is generated atthe softmax layer of ConvNet 2 next to fully connected layerCross entropy function is used for the classification at thefinal layer of the ConvNet 2 Softmax layer of ConvNet isshown in Figure 6

242 Support Vector Machine (SVM) SVM is a supervisedlearning model that uses an optimization method to identifysupport vectors Si weights αi and bias b )e classification isbeing considered to classify vectors x according to thefollowing equation

Rotations Translations

Figure 4 Data augmentation process through rotations andtranslations of a cervical cell

(a) (b) (c) (d)

(e) (f ) (g)

Figure 3 Examples of (andashc) normal cells and (dndashg) abnormal cells (a) superficial (b) intermediate (c) columnar (d) mild (e) moderate (f )severe (g) carcinoma


c 1113944i

αik si x( 1113857 + b (3)

where k is the kernel function depending on the modelassumed for decision boundary In case of a linear kernel k isa dot product If cge 0 then x is classified as a member of thefirst group and otherwise the second group Error correctingcode classifier is trained using support vector machine )ebatch size is set to 256 )e training set is applied to the

classifier along with deep hierarchical feature vector usingConvNet 2 Validation data are used in SVM to calculatevalidation accuracy

243 GentleBoost Ensemble of Decision Tree (GEDT)Regression trees or GEDTare used to predict the response ofthe data )e classification decisions are made when thequery sample follows the path from the initial or root node tothe end or leaf node

In GEDT an ensemble of trees is used that is based onmajority voting It is trained on the training data numbers oftrees are set to 100 and batch size is set to 256 Validationaccuracy of the Herlev dataset is evaluated using validationdata

244 Aggregated Score Evaluation of the cervical cellclassification task is done using 5-fold cross validation on theHerlev dataset for both 2-class and 7-class problems )eperformance metrics used for evaluation include accuracyF1 score area under the curve specificity and sensitivityFinally the count of correct classification score is obtainedfor each cell from all the categories in the Herlev dataset

3 Experiments and Results

31 Experimental Protocol In the training stage the convand pool layers of Alexnet ie ConvNet 1 as shown inFigure 5 are used as initial layers for the ConvNet 2Random weights are initialized to fc In order to trainConvNet 2 a patch of size 227 times 227 is cropped randomlyfrom each augmented image to make trainingtest images

Pretraining

Transfer learning

Conv 1ImageNet dataset

Cervical cellsSR

SVM

GEDT

Output layerfc

fc

pool 1Conv 2 pool 2

Conv 3pool 3

Conv 1

Image patch extraction

Fine-tuningFeatures extractionand classificationpool 1

Conv 2 pool 2Conv 3pool 3

Figure 5 Procedure of feature learning and classification using convolutional neural networks

Table 2 Configuration of ConvNet 2

Filter size Channel Stride PaddingInput mdash 3 mdash mdash(conv1) 11times 11 96 4 mdash(pool1) 3times 3 96 2 mdash(conv2) 5times 5 256 1 2(pool2) 3times 3 256 2 mdash(conv3) 3times 3 384 1 1(conv4) 3times 3 384 1 1(pool5) 3times 3 256 2 mdash(fc6) mdash 4096 mdash mdash(fc7) mdash 4096 mdash mdash

(fc) mdash 7 (7 classes) mdash mdash2 (2 classes)

Fully connectioncomponent

Somax activationfunction

Cross-entropyloss

Somax layer

Featuresinput

Figure 6 Softmax layer of ConvNet 2


compatible to the input nodes of the network To achievezero-center normalization a mean image over the dataset issubtracted Stochastic gradient descent (SGD) is used fortraining ConvNet 2 using 30 epochs Small batches of imagepatches are fed to ConvNet 2 and validation accuracy ofbatches is evaluated )e size of mini batch is set to 256Initial learning rate for convolutional and pooling layers isset to 00001 which is set to decrease with a factor of 10 afterevery 10 epochs L2 regularization and momentum can betuned to reduce overfitting and speed up learning process ofthe ConvNet 2 [25] L2 regularization or weight decay andmomentum are empirically set to 00005 and 09 respec-tively Finally the network is trained using a randomlyselected subset of epochs and validated for its accuracy )emodel having a minimum validation error is used forclassification application

In order to test the system against an unseen imagemultiple cropped patches of test images each having singlecell are generated from the original images containingmultiple cells Abnormal score of each crop is generated bythe classifier )e abnormal scores of all (Ntest times Ncrop)

patches of the test image are aggregated to generate the finalscore [47] Patches of test image Ntest 300 (10 rotations times

30 translations) are generated same as for training imagesFurthermore ten cropped images (Ntest) (four cornercenter of cell ie nucleus portion and their mirroredimages) are generated from each of test patch)ese (Ntest times

Ncrop) image patches are input to the classifier )e pre-diction score of the classifiers (SR SVM and GEDT) is thenaggregated to calculate the final score as shown in Figure 7

32 Experimental Results and Evaluation

321 ConvNet 2 Learning Results ConvNet 2 is fine-tunedon the Herlev dataset for 2-class and 7-class problems using30 epochs It is observed that after 10 epochs the validationaccuracy reaches its maximum value ie 09935 for the 2-class problem and 08745 for the 7-class problem Figure 8illustrates a fine-tuning process of ConvNet 2 during 30training epochs

)ese results are improved by considering variousclassifiers GEDT provides better performance with refer-ence to both classes because it exploits the randomness ofdata more efficiently as compared to other classifiers Table 3shows the comparison of SR SVM and GEDT

)e structure of the layers of network is also beingobserved after passing a test image to the fine-tuned net-work Features learned at the first layer ie (conv1) aremore generic as shown in Figure 9

It can be seen that these learned filters contain gradientsof different frequencies blobs and orientations of colors Aswe go deeper in convolutional layers (conv1 minus conv5) thefeatures become more prominent and provide more in-formation Figure 10 shows the feature learning results in(conv2 minus conv5) for a test cervical cell image

)e strongest activation is also shown in Figure 11 at thepool layer )e white pixels show strong positive activationwhile black pixels provide strong negative activation and

gray does not activate strongly It is also observed that thestrongest activation initiates negatively on right edges andpositively on left edges

)e feature set at fully connected layer is also exploredand it is observed that features are more abstract as com-pared to the previous layers as shown in Figure 12 Fullyconnected layer provides features learned for 7 classes

322 ConvNet 2 Testing Results It is a common knowledgethat a single evaluation metric is not appropriate to evaluatethe performance of a given algorithm due to the presence ofsome imbalanced classes in the dataset or a large number oftraining labels [48] )erefore the performance of the deepmodel is reported in terms of four distinct metrics includingaccuracy sensitivity specificity and F1 score as proposed inthe previous studies [49] )ese performance parameters arecalculated using the following equations

accuracy TP + TN

TP + FP + TN + FN

sensitivity TP

TP + FN

specificity TN

TN + FP

F1 2precisionrecallprecision + recall

(4)

where the precision and recall are expressed as

precision TP

TP + FP

recall TP

TP + FN

(5)

In the above equations true positive (TP) is defined as thenumber of malignant cell images classified as malignant andtrue negative (TN) is the number of benign cell imagesclassified as benign False positive (FP) is the number ofbenign cell images identified as malignant cell images andfalse negative (FN) is the number of malignant cell imagesclassified as benign

In ConvNet 2 a test set of cervical images using themultiple crop testing scheme is considered with threeclassifiers ie SR SVM and GEDT It can be seen thatGEDT again outperforms the classification accuracy of SRand SVM in test results also )e results are presented inTable 4

By analyzing the class-wise accuracy one can observethat the proposedmethod can predict the cervical cell imageswell )e classification accuracy of each of the seven cellcategories is calculated by feeding all the images as test to theclassifiers It is observed that GEDT shows superior per-formance on class 1 class 2 class 4 and class 5 because of itsability to eliminate irrelevant features and to extract decisionrules from decision trees efficiently )e performanceslightly deteriorates for class 3 and class 6 because theirfeatures are very close to each other causing confusion )e


1009080706050

Accu

racy

40302010

0 5 10 15Epoch

20 25 30

Val accuracy (2 class)Val accuracy (7 class)

Figure 8 )e fine-tuning process of ConvNet 2 during 30 epochs for 2-class and 7-class

Table 3 )e validation accuracy comparison of three classifiers

Validation accuracyClasses SR SVM GEDT2-class problem 9935 998 1007-class problem 8745 9620 9927

(a) (b)

Figure 9 (a) Test image (b) feature learned at (conv1)

C1

C2

C3

C4

C5

Multiple crops

Test image

C5 C4 C3 C2 C1Classifier Score

Aggregatedscore

Figure 7 Multiple crop testing scheme


classification accuracy of class 3 and class 6 is 9750 and9779 respectively Classification accuracy for class 7 is9920 )e average accuracy achieved by GEDT for un-derlying task is 9921 )ese results are illustrated inFigure 13

)e evaluation parameters of the classification perfor-mance ie accuracy F1 score area under the curvespecificity and sensitivity of the trained ConvNet 2 aredisplayed in Tables 5 and 6 where the performance

comparison of proposed work with [1339] and [50ndash56] ispresented We have proposed two scenarios with differentclassifiers ie SVM and GEDT )e mean values of accu-racy F1 score area under the curve (AUC) sensitivityand specificity of fine-tuned ConvNet 2 with GEDTclassifier are 996 9914 999 9930 and 9935respectively for the 2-class problem )ese are 98859877 998 988 and 9974 respectively for the 7-class problem

(a) (b)

(c) (d)

Figure 10 Features maps (a)(conv2) (b)(conv3) (c)(conv4) (d) Conv5

Figure 11 Strongest activation at the deepest (pool) layer Figure 12 Fully connected layer feature maps for 7 classes


)e accuracy of our system ie ConvNet with GEDT is996 for 2-class and 9885 for 7-class compared to 995and 912 for [39] respectively )is indicates that theprediction accuracy of our classification model is better thanthe existing models Similarly the sensitivity of 9938 and9930 implies better performance of the proposed methodcompared to the existing methods in classifying the cervicalcell images Likewise the values of specificity and accuracy ofproposed system for the 2-class problem are better thanprevious methods in [15 16 34 36 38 41 56]

)e images of cervical cells that are correctly classified ormisclassified are also analyzed Figure 14 shows the correctlyclassified malignant cell images columns 1 to 4 are milddysplasia moderate dysplasia severe dysplasia and carci-noma respectively

Figure 15 shows the result for test cell images that aremisclassified (normal misclassified as abnormal and ab-normal misclassified as normal)

323 Computational Complexity In the training phaseConvNet 1 is trained on the Corei7 machine with clockspeed 28 GHz NVidia 1080Ti GPU and 8GB ofmemory onMATLAB R2017b )e average training time for ConvNet 2running for 30 epochs is about 4 hours and 30minutes forthe 2-class and 8 hours and 20minutes for 7-class problemIn the testing phase the system takes 8 seconds to classify atest image into normal and abnormal classes Using multiplecrop testing ie (Npatches times Ncrops 3000) classificationsand score aggregation the average time for the testing of onecell image is around 8 seconds

4 Results-Related Discussion

)e experimental results presented in this study suggest thefollowing key observations

(1) Compared with the traditional prior feature ex-traction schemes the proposed cervical cellscreening system is more effective and robust )is isbecause ConvNets have been used to encode cervicalcancer specific features automatically In the tradi-tional methods such as [11 12 14 15] cervical cellsextraction strategies are hand-crafted which limit thesuccess of overall system Moreover in the presenceof large irregular shapes appearance dissimilaritiesand cervical cell clusters between malignant andbenign cell nucleus are the major problems in ac-curately segmenting the cytoplasm and nucleus Incontrast this method uses automatic cervical cellfeatures extraction to encode cancerous cell repre-sentation and thus achieve superior classificationaccuracy across a range of cells severity

(2) In order to prevent overfitting a data augmentationtechnique suitable for the underlying task of cervicalcell grading has been proposed )e training andvalidation losses for 30 epochs have been evaluatedto analyze the impact of the proposed data aug-mentation on classification accuracy It is observedthat the rate of overfitting is greatly reduced whenthe data augmentation strategy is applied to train ourclassification model )e smaller difference betweentraining and validation losses caused by data aug-mentation is presented in Figure 16 It indicates thathow this approach is fruitful for the classificationmodel to learn the most discriminative features forthe desired task Furthermore the proposed modelworks across a variety of cervical cells and preservesthe discriminative information during trainingWhile in the testing stage a cell image with arbitrarylevel of severity can be easily classified into the truegrading level Hence this suggests the efficacy of ourmethod to avoid the classification model fromoverfitting and shows robustness for classificationaccuracy against varying nature of cervical cells

(3) A multiple crop testing scheme is also used withthree classifiers to calculate the accuracy of all in-dividual classes of cervical cell images )e class-wiseaccuracy displayed in Figure 13 shows if the cervicalcells are more clear the classification ability of oursystem is more robust For example classificationaccuracy for class 1 class 2 class 4 and class 5 is thehighest ie 100 among all other classes It is be-cause this type of cells can be identified more ef-fectively by the underlying model )e classificationaccuracy for class 7 is 9920 Conversely theclassification performance slightly degrades for theclass 3 (9750) and class 6 (9779) because theirfeatures are very close to each other owing to lessercervical cells specific discriminative informationpresented to the model

(4) In a general way a single performance metric canlead to inappropriate classification results due tosome imbalanced classes in the dataset or too smallor large number of training samples )e literature

Table 4 Accuracy after classifier trained on deep features

AccuracyClasses SR SVM GEDT2-class problem 9980 9950 99607-class problem 9721 9812 9885

100 100 100 100

9779

992

975

1 2 3 4 5

Class-wise accuracy

6 7

Accuracy ()

Figure 13 Accuracy distributions according to all 7 classes


review of the existingmethods on cervical cell imagessuch as [13 39] and [50ndash56] shows classificationperformance in terms of accuracy metric only Onthe contrary we have considered four distinct

evaluation metrics including accuracy sensitivityspecificity and area under the curve )e experi-mental results displayed in Tables 5 and 6 show theconsistent performance of our proposed models in

Table 6 Comparison of the proposed work with existing state-of-the-art using the Herlev dataset

Systems Features Sensitivity () Specificity () Accuracy ()[34] Pap smear benchmarks 988 793 936[56] Genetic algorithm-based feature 985 921 968[16] PCA with SVM mdash mdash 969 (F1 score)[41] Local binary pattern mdash mdash 964 (AUC)[38] Ensemble classifier 990 897 965[36] CNN 982 983 983[15] Enhanced fuzzy C-means 9928 9747 9888Proposed ConvNet with SVM 9938 9920 995Proposed ConvNet with GEDT 9930 9935 996)e performance metrics are evaluated for the 2-class problem

Figure 14 Correctly classified malignant cell images

(a) (b)

Figure 15 Cervical cell images that are misclassified (a) normal as abnormal (b) abnormal as normal

Table 5 Accuracy comparison of the systems for 2-class and 7-class problems

Systems Features Dataset Accuracy[50] SVM 149 images of Herlev 98 (2-class)[51] k-NNs and ANNs Herlev k-NNs 88 ANNs 54 (2-class)[52] 20 features fuzzy C-means 2nd order NN Herlev 984 (2-class)[53] Decision tree Local 675 (2-class)[54] Backpropagation neural networks Local 956 (2-class) 7806 (7-class)[55] 20 features C-meansfuzzy clustering Herlev 94ndash96 (2-class) 72ndash80 (7-class)[13] Shape texture Gabor SVM Local 96 (2-class)[56] 20 features GA Herlev 98 (2-class) 969 (7-class)[39] CNN-ELM-based system Herlev 995 (2-class) 912 (7-class)Proposed ConvNet with SVM Herlev 995 (2-class) 9812 (7-class)Proposed ConvNet with GEDT Herlev 996 (2-class) 9885 (7-class)


cervical cell images classification across differentevaluation metrics We have proposed two scenarioswith different classifiers ie SVM and GEDT It isnoted that the proposed scheme outperforms allprevious approaches )is is despite of the fact thattraining and test images also contained images ofoverlapping cells )is exceptional performance ismainly because of following reasons (1) during thetraining stage transfer learning is used and thenetwork is finally fine-tuned on the Herlev dataset(2) the trained network is used only for extraction ofdeep features and (3) the extracted features are thenfed to more robust classifiers like SVM and GEDTwhich are used for final classification )is suggeststhe effectiveness of ourmethod for underlying task inthe presence of a wide variety of cervical cell imagesranging from class 1 to class 7

(5) )e structure of different layers of fine-tuned net-work is also explored It is seen that the featureslearned in the initial conv layers are more general-ized and as we move deeper into the network theextracted features tend to become more abstract )efeatures learned at fully connected layers are dis-played in Figures 9ndash12

(6) Despite higher performance of deep learning-basedcervical cell screening system it has some limita-tions Classification time of testing a cropped singlecell image is 8 seconds for the system that is very slowin clinical setting as there are large numbers ofsamples from one PAP smear slide )is limitationcan be addressed by neglecting the process of dataaugmentation step for the test data and only mul-tiple crop testing can be used for classificationproblem )is increases the speed of the system as itrequires only 008 seconds for classification but theaccuracy of the system is compromised by 15Although classification accuracy of the system on the

Herlev dataset is high there is room for furtherimprovement

5 Conclusions and Future Work

)is paper proposes an automatic cervical cancer screeningsystem using convolutional neural network Unlike previousmethods which are based upon cytoplasmnucleus seg-mentation and hand-crafted features our method auto-matically extracts deep features embedded in the cell imagepatch for classification )is system requires cells withcoarsely centered nucleus as the network input Transferlearning is used for pretraining and initial weights or featuremaps are transferred from a pretrained network to a newconvolutional neural network for fine-tuning on the cervicalcell dataset )e features learned by the new fine-tunednetwork are extracted and given as input to different clas-sifiers ie SR SVM and GEDT)e validation results for 2-class and 7-class problems are analyzed To test a single cellimage different test image patches are generated same astraining data and the multiple crop testing scheme has beencarried out on all patches to achieve classifier score It isfurther aggregated for the final score )e proposed methodyields the highest performance as compared to previousstate-of-the-art approaches in terms of classification accu-racy sensitivity specificity F1 score and area under thecurve on the Herlev Pap smear dataset It is anticipated that asegmentation free highly accurate cervical cell classificationsystem of this type is a promising approach for the devel-opment of auto-assisted cervical cancer screening system

In future the effect of system on field of view imagescontaining overlapping cells is to be analyzed )e systemshould avoid the misclassification of overlapping objectsSpecific classifiers relying on deep learning may be used tocater these problems Moreover deep learning-based cer-vical cell classification still needs to be explored for high-precision diagnosis

25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30

Val loss (2 class)Val loss (7 class)

Train loss (2 class)Train loss (7 class)

Figure 16 Effect of data augmentation on training and validation loss for 2-class and 7-class


Data Availability

)e data used to support the findings of this study will bemade available on request

Conflicts of Interest

)e authors declare that there are no conflicts of interestregarding the publication of this paper

References

[1] C Marth F Landoni S Mahner M McCormackA Gonzalez-Martin and N Colombo ldquoCervical cancerESMO clinical practice guidelines for diagnosis treatmentand follow-uprdquo Annals of Oncology vol 28 p iv72 2017

[2] M F Janicek and H E Averette ldquoCervical cancer preventiondiagnosis and therapeuticsrdquo CA A Cancer Journal for Cli-nicians vol 51 no 2 p 92 2001

[3] R M Demay ldquoCommon problems in papanicolaou smearinterpretationrdquo Archives of Pathology and Laboratory Medi-cine vol 121 no 3 pp 229ndash238 1997

[4] M Arya N Mittal and G Singh ldquoCervical cancer detectionusing segmentation on pap smear imagesrdquo Proceedings of theInternational Conference on Informatics and Analytics vol 162016

[5] F Bray B Carstensen H Moslashller et al ldquoIncidence trends ofadenocarcinoma of the cervix in 13 European countriesrdquoCancer Epidemiology Biomarkers amp Prevention vol 14 no 9pp 2191ndash2199 2005

[6] D Saslow D Solomon H W Lawson et al ldquoAmericanCancer Society American Society for Colposcopy and Cer-vical Pathology and American Society for Clinical Pathologyscreening guidelines for the prevention and early detection ofcervical cancerrdquo CA A Cancer Journal for Clinicians vol 62no 3 2012

[7] G G Birdsong ldquoAutomated screening of cervical cytologyspecimensrdquo Human Pathology vol 27 no 5 pp 468ndash4811996

[8] H C Kitchener R Blanks G Dunn et al ldquoAutomation-assisted versus manual reading of cervical cytology(MAVARIC) a randomised controlled trialrdquo Be LancetOncology vol 12 no 1 pp 56ndash64 2011

[9] E Bengtsson PMalm E Bengtsson and PMalm ldquoScreeningfor cervical cancer using automated analysis of PAP-smearsrdquoComputational and Mathematical Methods in Medicinevol 2014 p 1 2014

[10] L Zhang H Kong C Ting Chin et al ldquoAutomation-assistedcervical cancer screening in manual liquid-based cytologywith hematoxylin and eosin stainingrdquo Cytometry Part Avol 85 no 3 p 214 2014

[11] E Bak K van Najarian and J P Brockway ldquoEfficient seg-mentation framework of cell images in noise environmentsrdquoin Proceedings of Engineering in Medicine and Biology Society2004 IEMBSrsquo04 26th Annual International Conference of theIEEE National Center for Biotechnology InformationBethesda MD USA July 2004

[12] Rahmadwati G Naghdy M Ros C Todd andE Norahmawati ldquoCervical cancer classification using Gaborfiltersrdquo in Proceedings of 2011 1st IEEE International Con-ference on Healthcare Informatics Imaging and SystemsBiology July 2011

[13] P Wang L Wang Y Li Q Song S Lv and X Hu ldquoAu-tomatic cell nuclei segmentation and classification of cervical

pap smear imagesrdquo Biomedical Signal Processing and Controlvol 48 pp 93ndash103 2019

[14] T Chankong N )eera-Umpon and S AuephanwiriyakulldquoAutomatic cervical cell segmentation and classification inPap smearsrdquo Computer Methods and Programs in Biomedi-cine vol 113 no 2 p 539 2014

[15] W William A Ware A H Basaza-Ejiri and J ObungolochldquoA pap-smear analysis tool (PAT) for detection of cervicalcancer from pap-smear imagesrdquo BioMedical EngineeringOnline vol 18 no 1 pp 1ndash22 2019

[16] M E Plissiti C Nikou and A Charchanti ldquoAutomateddetection of cell nuclei in pap smear images using morpho-logical reconstruction and clusteringrdquo IEEE Transactions onInformation Technology in Biomedicine vol 15 no 2pp 233ndash241 2011

[17] A Garrido and N Perez De La Blanca ldquoApplying deformabletemplates for cell image segmentationrdquo Pattern Recognitionvol 33 no 5 pp 821ndash832 2000

[18] R Saha M Bajger and G Lee ldquoPrior guided segmentationand nuclei feature based abnormality detection in cervicalcellsrdquo in Proceedings - 2019 IEEE 19th International Con-ference on Bioinformatics and Bioengineering BIBE 2019pp 742ndash746 Institute of Electrical and Electronics EngineersInc Athens Greece October2019

[19] N Lavanya Devi and P)irumurugan ldquoAutomated detectionof cervical cancerrdquo International Journal of InnovativeTechnology and Exploring Engineering vol 8 pp 1399ndash14012019

[20] M Jones and L R McNally ldquoA new approach for automatedimage segmentation of organs at risk in cervical cancerrdquoRadiology Imaging Cancer vol 2 no 2 2020

[21] Y Bar I Diamant L Wolf S Lieberman E Konen andH Greenspan ldquoChest pathology detection using deeplearning with non-medical trainingrdquo in Proceedings of 2015IEEE 12th International Symposium on Biomedical Imaging(ISBI) April 2015

[22] R Girshick J Donahue T Darrell and J Malik ldquoRegion-based convolutional networks for accurate object detectionand segmentationrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 38 no 1 2016

[23] Y LeCun Y Bengio and G Hinton ldquoDeep learningrdquoNaturevol 521 no 7553 p 436 2015

[24] Y LeCun B Boser J S Denker et al ldquoBackpropagationapplied to handwritten zip code recognitionrdquo Neural Com-putation vol 1 no 4 p 541 1989

[25] A Krizhevsky I Sutskever and G E Hinton ldquoImageNetclassification with deep convolutional neural networksrdquoAdvances In Neural Information Processing Systems vol 25no 2 2012

[26] H Greenspan B van Ginneken and R M Summers ldquoGuesteditorial deep learning in medical imaging overview andfuture promise of an exciting new techniquerdquo IEEE Trans-actions on Medical Imaging vol 35 no 5 pp 1153ndash11592016

[27] R M Summers ldquoProgress in fully automated abdominal CTinterpretationrdquo American Journal of Roentgenology vol 207no 1 Article ID 15334406 2016

[28] H-C Shin H R Roth M Gao et al ldquoDeep convolutionalneural networks for computer-aided detection CNN archi-tectures dataset characteristics and transfer learningrdquo IEEETransactions on Medical Imaging vol 35 no 5 p 1285 2016

[29] H R Roth L Lu J Liu et al ldquoImproving computer-aideddetection using_newline convolutional neural networks and


random view aggregationrdquo IEEE Transactions on MedicalImaging vol 35 no 5 p 1170 2016

[30] T Xu H Zhang C Xin et al ldquoMulti-feature based bench-mark for cervical dysplasia classification evaluationrdquo PatternRecognition vol 63 pp 468ndash475 2017

[31] T Xu H Zhang X Huang S Zhang and D N MetaxasldquoMultimodal deep learning for cervical dysplasia diagnosisrdquoLecture Notes in Computer Science (Including Subseries LectureNotes in Artificial Intelligence and Lecture Notes in Bio-informatics) vol 9901 pp 115ndash123 2016

[32] P Buyssens A Elmoataz and O Lezoray ldquoMultiscale con-volutional neural networks for vision-based classification ofcellsrdquo Lecture Notes in Computer Science (Including SubseriesLecture Notes in Artificial Intelligence and Lecture Notes inBioinformatics) vol 7725 pp 342ndash352 2013

[33] J Lemley S Bazrafkan and P Corcoran ldquoSmart augmen-tation learning an optimal data augmentation strategyrdquo IEEEAccess vol 5 p 5858 2017

[34] J Jantzen J Norup G Dounias and B Bjerregaard ldquoPap-smear benchmark data for pattern classificationrdquo in Pro-ceedings of NiSIS 2005 Nature Inspired Smart InformationSystems January 2005

[35] J Yosinski J Clune Y Bengio and H Lipson ldquoHowtransferable are features in deep neural networks rdquo 2014httpsarxivorgabs14111792

[36] L Zhang L Le Lu I Nogues R M Summers S Liu andJ Yao ldquoDeepPap deep convolutional networks for cervicalcell classificationrdquo IEEE Journal of Biomedical and HealthInformatics vol 21 no 6 p 1633 2017

[37] C Li D Xue X Zhou et al ldquoTransfer learning based clas-sification of cervical cancer immunohistochemistry imagesrdquoin Proceedings of the Bird International Symposium on ImageComputing and Digital Medicine - ISICDM 2019 pp 102ndash106ACM Xirsquoan China August 2019

[38] K Bora M Chowdhury L B Mahanta M K Kundu andA K Das ldquoAutomated classification of pap smear images todetect cervical dysplasiardquo Computer Methods and Programs inBiomedicine vol 138 pp 31ndash47 2017

[39] A Ghoneim G Muhammad and M S Hossain ldquoCervicalcancer classification using convolutional neural networks andextreme learning machinesrdquo Future Generation ComputerSystems vol 102 pp 643ndash649 2020

[40] L Nanni A Lumini and S Brahnam ldquoLocal binary patternsvariants as texture descriptors for medical image analysisrdquoArtificial Intelligence in Medicine vol 49 no 2 pp 117ndash1252010

[41] Y Guo G Zhao andM Pietikainen ldquoDiscriminative featuresfor texture descriptionrdquo Pattern Recognition vol 45 no 10pp 3834ndash3843 2012

[42] B Sokouti S Haghipour and A D Tabrizi ldquoA framework fordiagnosing cervical cancer disease based on feedforward MLPneural network and )inPrep histopathological cell imagefeaturesrdquo Neural Computing and Applications vol 24 no 1p 221 2014

[43] Z Gao L Wang L Zhou and J Zhang ldquoHEp-2 cell imageclassification with deep convolutional neural networksrdquo IEEEJournal of Biomedical and Health Informatics vol 21 no 2p 416 2017

[44] Z Lu G Carneiro A P Bradley et al ldquoEvaluation of threealgorithms for the segmentation of overlapping cervical cellsrdquoIEEE Journal of Biomedical and Health Informatics vol 21no 2 p 441 2017

[45] H He and E A Garcia ldquoLearning from imbalanced datardquoIEEE Transactions on Knowledge and Data Engineeringvol 21 no 9 2009

[46] D Jia W Dong R Socher L Li-Jia L Kai and L Fei-FeildquoImageNet a large-scale hierarchical image databaserdquo inProceedings of 2009 IEEE Conference on Computer Vision andPattern Recognition June 2009

[47] O Russakovsky J Deng H Su et al ldquoImageNet large scalevisual recognition challengerdquo International Journal of Com-puter Vision vol 115 no 3 p 211 2015

[48] A Moayedikia K L Ong Y L Boo W G Yeoh andR Jensen ldquoFeature selection for high dimensional imbalancedclass data using harmony searchrdquo Engineering Applications ofArtificial Intelligence vol 57 pp 38ndash49 2017

[49] D Powers ldquoEvaluation from precision recall and F-measureto roc informedness markedness amp correlationrdquo Journal ofMachine Learning Technologies vol 35 no 1 pp 41ndash64 2011

[50] J Zhang and Y Liu ldquoCervical cancer detection using SVMbased feature screeningrdquo International Conference on MedicalImage Computing and Computer-Assisted Interventionvol 3217 pp 873ndash880 2004

[51] P K Malli and S Nandyal ldquoMachine learning technique fordetection of cervical cancer using k-NN and artificial neuralnetworkrdquo International Journal of Emerging Trends amp Tech-nology in Computer Science (IJETTCS) vol 6 2017

[52] N Ampazis G Dounias and J Jantzen ldquoPap-smear classi-fication using efficient second order neural network trainingalgorithmsrdquo in Proceeding of the Hellenic Conference onArtificial Intelligence pp 230ndash245 Samos Greece May 2004

[53] S Sharma ldquoCervical cancer stage prediction using decisiontree approach of machine learningrdquo International Journal ofAdvanced Research in Computer and Communication Engi-neering vol 5 pp 345ndash348 2016

[54] R Gupta A Sarwar and V Sharma ldquoScreening of cervicalcancer by artificial intelligence based analysis of digitizedpapanicolaou-smear imagesrdquo International Journal of Con-temporary Medical Research vol 4 2017

[55] G Dounias B Bjerregaard J Jantzen et al ldquoAutomatedidentification of cancerous smears using various competitiveintelligent techniquesrdquo Oncology Reports vol 15 pp 1001ndash1006 2006

[56] Y Marinakis G Dounias and J Jantzen ldquoPap smear diag-nosis using a hybrid intelligent scheme focusing on geneticalgorithm based feature selection and nearest neighborclassificationrdquo Computers in Biology and Medicine vol 39no 1 p 69 2009


precancerous and cancerous cells In every 10 to 15 positivecases there is a chance of one case to be missed in con-ventional screening [3]

)e rate of the incidence of cervical cancer is lesser in theUSA and other parts of developed countries because of earlydetection and better screening methods [4] Its rate of oc-currence has been dropped by 80 since the screeningsystems are introduced in some Nordic countries In Sweden[5] it is dropped by 65 during the last four decades and theoccurrence of cervical cancer and mortality figures are stableover the last decade However improved screening systemsare still unavailable in underdeveloped countries partly dueto the complexity and tedious nature of manual screening ofabnormal cells from a cervical cytology specimen [6 7]While auto-assisted mass screening techniques can boostefficiency they are not accurate enough to be used as aprimary tool for cervical screening [8]

During the past few years extensive research has beencarried out for the development of computer-assisted au-tomated reading systems based on cell image analysis[7 9 10])emanual screening process is normally initiatedwith the collection of cervical cell samples from the uterinecervix and their placement on a glass slide After visualinspection under a microscope these are classified intodifferent categories)e shape size texture ratio of nucleusand cytoplasm are the main characteristics for the classifi-cation task Hence for an automated system the first stepmay include segmentation of images of cell samples toextract regions of interest containing single cells with nu-cleus and cytoplasm from the noncell regions )is initialsegmentation is then followed by separation of main cellcomponents including the nucleus and cytoplasm and ex-traction of their shapetextural features However theseparation of main components and shape feature extractionis not an integral part of an automated screening system asproposed schemes in the literature include both options iewith geometrical feature extraction and without priorextraction

For a system that includes prior feature extraction ac-curate segmentation of nucleus from cytoplasm in cervicalcell images is a difficult task and is prone to error thus

limiting the success of overall system )e presence of largeirregular shapes appearance dissimilarities and cell clustersbetween malignant and benign cell nucleus is the majorproblem in accurately segmenting the cytoplasm and nu-cleus Various segmentation algorithms have been proposedby researchers to segment out cell components An iterativealgorithm for assigning pixels based on a statistical criterionfunction was proposed in [11] to separate the nucleus cy-toplasm and background In another study [12 13] Gaborfilters were applied for exploiting textural variation of thecervical cells to segment out regions of interest FuzzyC-means clustering was used in [14 15] to segment thesingle cell images into nucleus cytoplasm and backgroundHowever if the overlapping cells are taken into account theclassification accuracy is decreased significantly)erefore amajority of the presented segmentation approaches [16ndash20][11 12 14] are effective in terms of their performance forsingle and clear cervical cell images only but in the case ofoverlapping cells or other shape changes they lack theperformance accuracy

To overcome this dependency on segmentation manytechniques have been proposed during the past few yearswhich do not include prior segmentation and directlyclassify the unsegmented cell images A pixel-level classifi-cation method is proposed in [21] to classify normal andabnormal cells without prior segmentation using block-wisefeature selection and extraction techniques However thevalidation accuracy of the proposed algorithm is not up tothe mark In [22] block image processing was proposed thatincludes cropping arbitrary image blocks prior to featureextraction and the cropped blocks are then classified usingSVM However in their approach arbitrary cropping couldpotentially separate a full cell into distinct patches

Recently feature representation in image classificationproblems based on deep learning methods has become morepopular [23] In particular convolutional neural networks(ConvNets) [24] have achieved unprecedented results in the2012 ImageNet Large Scale Visual Recognition Challengewhich consisted of classifying natural images in theImageNet dataset into 1000 fine-grained categories [25]Besides they have drastically increased the accuracy in thefield of medical imaging [26 27] specifically classification oflung diseases and lymph nodes in CT images [28 29] anddetecting cervical intraepithelial neoplasia based on cervi-gram images [30] or multimodal data [31] ConvNets havealso shown superior performance in the classification of cellimages for diagnosis of pleural cancer [32]

However large datasets are essential to achieve highperformance and to overcome the problem of overfittingwith ConvNets [33] )is is a major limitation in applyingConvNets to the cervical cell classification problem as in thecase of cervical cells and a limited number of annotateddatasets are available For instance the Herlev dataset [34]only contains 917 cervical cells with 675 abnormal and 242normal cells that are insufficient for ConvNets To overcomethis limitation recently image data augmentation tech-niques have been proposed to virtually increase the size oftraining datasets and reduce the problem of overfitting [25]Data augmentation can be achieved by linear transformation

Uterus Fallopian tube

Uterine lining

Vagina

CervixOvary

Figure 1 Female reproductive system Initially the cervical cancerstarts from the cervix the lower end of uterus






















ImageNetdatabase

Herlevdatabase

Trai

ning

dat

aTe

sting

dat

a

Testimage

Pre-processeddata

Pre-processing


Deep features

Aggregatedscore


Transfer learning















P y jz(i)

1113872 1113873 ϕsoftmax z(i)

1113872 1113873 e

z(i)

1113936kj0 e

z(i)

k

(1)



i0wixi w

Tx (2)





(a) (b) (c) (d)

(e) (f ) (g)



c 1113944i

αik si x( 1113857 + b (3)








Pretraining

Transfer learning


Cervical cellsSR

SVM

GEDT

Output layerfc

fc

pool 1Conv 2 pool 2

Conv 3pool 3

Conv 1










Cross-entropyloss

Somax layer

Featuresinput

















accuracy TP + TN

TP + FP + TN + FN

sensitivity TP

TP + FN

specificity TN

TN + FP


(4)


precision TP

TP + FP

recall TP

TP + FN

(5)





1009080706050

Accu

racy

40302010

0 5 10 15Epoch

20 25 30





(a) (b)


C1

C2

C3

C4

C5

Multiple crops

Test image


Aggregatedscore






(a) (b)

(c) (d)
















100 100 100 100

9779

992

975

1 2 3 4 5

Class-wise accuracy

6 7

Accuracy ()








(a) (b)












25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References

















































































ImageNetdatabase

Herlevdatabase

Trai

ning

dat

aTe

sting

dat

a

Testimage

Pre-processeddata

Pre-processing


Deep features

Aggregatedscore


Transfer learning















P y jz(i)

1113872 1113873 ϕsoftmax z(i)

1113872 1113873 e

z(i)

1113936kj0 e

z(i)

k

(1)



i0wixi w

Tx (2)





(a) (b) (c) (d)

(e) (f ) (g)



c 1113944i

αik si x( 1113857 + b (3)








Pretraining

Transfer learning


Cervical cellsSR

SVM

GEDT

Output layerfc

fc

pool 1Conv 2 pool 2

Conv 3pool 3

Conv 1










Cross-entropyloss

Somax layer

Featuresinput

















accuracy TP + TN

TP + FP + TN + FN

sensitivity TP

TP + FN

specificity TN

TN + FP


(4)


precision TP

TP + FP

recall TP

TP + FN

(5)





1009080706050

Accu

racy

40302010

0 5 10 15Epoch

20 25 30





(a) (b)


C1

C2

C3

C4

C5

Multiple crops

Test image


Aggregatedscore






(a) (b)

(c) (d)
















100 100 100 100

9779

992

975

1 2 3 4 5

Class-wise accuracy

6 7

Accuracy ()








(a) (b)












25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References


































































ImageNetdatabase

Herlevdatabase

Trai

ning

dat

aTe

sting

dat

a

Testimage

Pre-processeddata

Pre-processing


Deep features

Aggregatedscore


Transfer learning















P y jz(i)

1113872 1113873 ϕsoftmax z(i)

1113872 1113873 e

z(i)

1113936kj0 e

z(i)

k

(1)



i0wixi w

Tx (2)





(a) (b) (c) (d)

(e) (f ) (g)



c 1113944i

αik si x( 1113857 + b (3)








Pretraining

Transfer learning


Cervical cellsSR

SVM

GEDT

Output layerfc

fc

pool 1Conv 2 pool 2

Conv 3pool 3

Conv 1










Cross-entropyloss

Somax layer

Featuresinput

















accuracy TP + TN

TP + FP + TN + FN

sensitivity TP

TP + FN

specificity TN

TN + FP


(4)


precision TP

TP + FP

recall TP

TP + FN

(5)





1009080706050

Accu

racy

40302010

0 5 10 15Epoch

20 25 30





(a) (b)


C1

C2

C3

C4

C5

Multiple crops

Test image


Aggregatedscore






(a) (b)

(c) (d)
















100 100 100 100

9779

992

975

1 2 3 4 5

Class-wise accuracy

6 7

Accuracy ()








(a) (b)












25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References


































































P y jz(i)

1113872 1113873 ϕsoftmax z(i)

1113872 1113873 e

z(i)

1113936kj0 e

z(i)

k

(1)



i0wixi w

Tx (2)





(a) (b) (c) (d)

(e) (f ) (g)



c 1113944i

αik si x( 1113857 + b (3)








Pretraining

Transfer learning


Cervical cellsSR

SVM

GEDT

Output layerfc

fc

pool 1Conv 2 pool 2

Conv 3pool 3

Conv 1










Cross-entropyloss

Somax layer

Featuresinput

















accuracy TP + TN

TP + FP + TN + FN

sensitivity TP

TP + FN

specificity TN

TN + FP


(4)


precision TP

TP + FP

recall TP

TP + FN

(5)





1009080706050

Accu

racy

40302010

0 5 10 15Epoch

20 25 30





(a) (b)


C1

C2

C3

C4

C5

Multiple crops

Test image


Aggregatedscore






(a) (b)

(c) (d)
















100 100 100 100

9779

992

975

1 2 3 4 5

Class-wise accuracy

6 7

Accuracy ()








(a) (b)












25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References





























































c 1113944i

αik si x( 1113857 + b (3)








Pretraining

Transfer learning


Cervical cellsSR

SVM

GEDT

Output layerfc

fc

pool 1Conv 2 pool 2

Conv 3pool 3

Conv 1










Cross-entropyloss

Somax layer

Featuresinput

















accuracy TP + TN

TP + FP + TN + FN

sensitivity TP

TP + FN

specificity TN

TN + FP


(4)


precision TP

TP + FP

recall TP

TP + FN

(5)





1009080706050

Accu

racy

40302010

0 5 10 15Epoch

20 25 30





(a) (b)


C1

C2

C3

C4

C5

Multiple crops

Test image


Aggregatedscore






(a) (b)

(c) (d)
















100 100 100 100

9779

992

975

1 2 3 4 5

Class-wise accuracy

6 7

Accuracy ()








(a) (b)












25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References











































































accuracy TP + TN

TP + FP + TN + FN

sensitivity TP

TP + FN

specificity TN

TN + FP


(4)


precision TP

TP + FP

recall TP

TP + FN

(5)





1009080706050

Accu

racy

40302010

0 5 10 15Epoch

20 25 30





(a) (b)


C1

C2

C3

C4

C5

Multiple crops

Test image


Aggregatedscore






(a) (b)

(c) (d)
















100 100 100 100

9779

992

975

1 2 3 4 5

Class-wise accuracy

6 7

Accuracy ()








(a) (b)












25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References





























































1009080706050

Accu

racy

40302010

0 5 10 15Epoch

20 25 30





(a) (b)


C1

C2

C3

C4

C5

Multiple crops

Test image


Aggregatedscore






(a) (b)

(c) (d)
















100 100 100 100

9779

992

975

1 2 3 4 5

Class-wise accuracy

6 7

Accuracy ()








(a) (b)












25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References
































































(a) (b)

(c) (d)
















100 100 100 100

9779

992

975

1 2 3 4 5

Class-wise accuracy

6 7

Accuracy ()








(a) (b)












25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References









































































100 100 100 100

9779

992

975

1 2 3 4 5

Class-wise accuracy

6 7

Accuracy ()








(a) (b)












25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References


































































(a) (b)












25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References




































































25

2

15

1

Loss

05

00 5 10 15

Epochs20 25 30





Data Availability




References





























































Data Availability




References


























































































Documents

AnAutomaticMassScreeningSystemforCervicalCancer ...single normal cell and 160 image patches from each ab-normal cell. is transformation yields relative normal distribution, as the