Improved Integrated Deep Model for Pap-Smear Cell Analysis

Somasundaram DEVARAJ; Nirmala MADIAN; Gnanasaravanan SUBRAMANIAM; Rithaniya CHELLAMUTHU; Muralitharan KRISHANAN

doi:10.21078/JSSI-2023-0087

PDF(724 KB)

Journal of Systems Science and Information ›› 2024, Vol. 12 ›› Issue (1) : 113-124. DOI: 10.21078/JSSI-2023-0087

Improved Integrated Deep Model for Pap-Smear Cell Analysis

Author information +

History +

Abstract

Cervical cancer is the fourth most common malignancy to strike a woman globally. If discovered early enough, it can be effectively treated. Although there is a chance of error owing to human error, the Pap smear is a good tool for first screening for cervical cancer. It also takes a lot of time and effort to complete. The aim of this study was to reduce the possibility of error by automating the process of classifying cervical cancer using Pap smear images. For the purpose of this study, dual convolution neural networks with LSTM were employed to classify images due to deep learning approaches inspire distinct features and powerful classifiers for many computer vision applications. The proposed deep learning model based on convolution neural networks (CNN) with the long short-term memory (LSTM) network is to learn features which give better recognition accuracy. The overall model is known as Smear-net. In which 'smear' indicates 'pap-smear cancer cells' and 'net' refers to neural network. The parameters such as, Accuracy, Precision, Recall, Accuracy, Sensitivity, and Specificity are used to validate the models. The proposed method provides the improved accuracy of 99.57 percentage for classification of the pap-smear cells. The proposed approaches demonstrate the effectiveness of our contributions by testing and comparing with the state-of-the-art techniques.

Key words

pap-smear / deep learning / convolution neural network / smear-net

Cite this article

EndNote

Ris (Procite)

Bibtex

Download Citations

Somasundaram DEVARAJ , Nirmala MADIAN , Gnanasaravanan SUBRAMANIAM , Rithaniya CHELLAMUTHU , Muralitharan KRISHANAN. Improved Integrated Deep Model for Pap-Smear Cell Analysis. Journal of Systems Science and Information, 2024, 12(1): 113-124 https://doi.org/10.21078/JSSI-2023-0087

1 Introduction

According to World Health Organization during 2018 about 570000 women diagnosed with cervical cancer and around 311000 women died due to this cancer. Cervical cancer has now become a potential life threatening cancer. The prevention processes of cervical cancer are HIV vaccination and proper screening and treatment of precancerous lesion. One of the primary tools for cervical cancer screening is Papanicolaou (Pap) smear test. The cervical cancer occurs in the cervix region and is analysed using smear cell. Smear cell can be touching or overlapping when viewed through microscope. Many researches on segmentation and separation algorithms are proposed to analyse the overlapping and touching smear cells.

Pap smear cell segmentation is a specialized area in medical image analysis and help in early detection of cancer. There are various advanced computational techniques for identifying cervical cancer. Pap smear is a primary screening procedure for understanding the smear cell from complex microscope. The smear analysis helps in detecting abnormalities with better precision and efficiency. The efficiency is based on the classification accuracy of smear cell. The cervical cancer can be categorized into various classes like normal, precancerous and cervical cancer based on morphological features. Classification accuracy helps in reducing the morbidity and mortality rate of cancer and also leads to early diagnosis of cancer.

In recent years deep learning a subset of artificial intelligence has emerged as a transformative solution in smear cell classification. Deep learning algorithms, particularly convolutional neural networks (CNNs) are recognised for automatically and accurately categorizing cervical cells based on their morphological characteristics. This technological advancement holds immense promise in revolutionizing cervical cancer screening, offering the potential for earlier and more precise identification of abnormalities, including precancerous and cancerous cells. The limitation in classification accuracy is based on selecting the appropriate features.

Pap smear cell has cytoplasm and nucleus. Nucleus segmentation is the process of separating the nucleus region from the papsmear cell. An approach based on the image colour information was proposed which is based on colour characterization. Both single and overlapping cell clusters can be handled using the approach^[1–5]. A fuzzy-based processing of smear cell is presented that reduces the over segmentation problems by detecting the nucleus border using the watershed transformation and classifying the nuclei using markers^[6–8].

A method to segment the nucleus is performed using an active shape model (ASM) and a physical based model. The vibration of a spring mass system is used to represent the shape of the nuclei in the training^[9–13]. A global and local technique for segmenting the cervical cell to separate the cytoplasm region from the background and to represent the overall scheme, a multi-way graph cut technique is adopted^[14–19]. As a local strategy for dividing up the nucleus region, the local adaptive graph cut is employed and separation is performed based on concave and convex points. The nucleus and cytoplasm region are enhanced using linear contrast enhancement approach while the noise is eliminated using a median filter^[20–22]. The maximum grey gradient difference approach is used to extract the nucleus contour.

Nucleus and cytoplasm extraction was performed using the snake method called radiating gradient vector flow (RGVF). The effort to categorise the cervical cell as normal or abnormal by the Bayes rule, which is based on two decision criteria^{[23, 24]}. Using the parametric Bhattacharya distance formulation, texture features are retrieved. SVM-based classifier is used to categorise multispectral Pap smear images. For the purpose of distinguishing cytoplasm and nuclei, a system with a dual wavelength technique is used. Segmentation, feature extraction, and smear categorization are the first three steps of the approach. For classification, linear discrimination analysis is utilised^[25-28].

The classification of Pap smear cell is based on numerical measurement and can be analysed by data analysis^[29-32]. Seven groups or classes are available like squamous cell carcinoma, mild, moderate, and severe non-keratinizing squamous epithelial dysplasia, columnar epithelial dysplasia, and superficial squamous epithelial. Artificial neural networks are used for classifying cervical cells. The features are used to classify the areas of the cytoplasm and nucleus^[33-36]. Six characteristics relate to the nucleus, whereas the remaining characteristics relate to the cytoplasm. A mask based regional convolutional neural network (Mask R-CNN) presented for the classification of smear cell classification^[37-39]. Still, in less developed nations, their influence on cervical cancer screening has been minimal. Considerable research has been done since the first computers were made in an attempt to replace or augment human visual evaluation of Pap smears with computer-aided analysis. But the issue proved to be trickier than anticipated. The examination of Pap smears with computer assistance has proven to be a challenging procedure. The methodology presented in this study can be used to a wide range of Pap smear analysis systems, but it is especially relevant to low-cost systems that are expected to have a substantial positive impact on emerging countries. The proposed smearnet is outperforming than traditional classification methods.

2 Proposed Method

The proposed model consists of two stages as shown in Figure 1: First, the pre-processing stage for handling illumination variance and other to enhance the edges of each image. Second, the pre-processed input is given separately to the dual CNN layer which learns the spatial features and generate feature maps. Finally, the feature maps obtained from two fully connected layers are fused. Further, the global average pooling (GAP) is used to reduce feature maps and the expression is predicted using SoftMax layer. The layer information of the Smearnet as shown in Table 1.

Figure 1 Overall pap-smear cell classification

Full size|PPT slide

Table 1 Smearnet Architecture

Layer	Filter	Stride	Size
Input	–	–	$96 \times 96$
Conv 1	32	1	$5 \times 5$
Maxpooling 1	–	2	$3 \times 3$
Conv 2	32	1	$5 \times 5$
Maxpooling 2	–	2	$3 \times 3$
Conv 3	64	1	$5 \times 5$
Maxpooling 3	–	2	$3 \times 3$
Conv 4	128	1	$5 \times 5$
Maxpooling 4	–	2	$3 \times 3$
Conv 5	256	1	$5 \times 5$
Maxpooling 5	–	2	$3 \times 3$
FC1	512	–	–
FC2	7	–	–

Data Pre-Processing: Consider a video-sequence input which is converted into 16 frames per second. Then the smear cell detection is employed to detect smear regions from the successive frames. The detected face is cropped and resized to

96 \times 96

-pixel size. The probability density function of the resized image is given as,

P (G_{k}) = \frac{N_{k}}{N},

(1)

where

G_{k}

is the greyscale image,

N_{k}

is the number of times the occurrence of

G_{k}

, and

N

is the total pixel. The overall process as shown in the Figure 2. The quality of image is improved with certain pre-processing techniques to learn and classify expression features effectively. The two pre-processing techniques are contrast enhancement technique, edge enhancement technique. Histogram equalization (HE) is one of the commonly used techniques for image contrast enhancement. The HE is simple and effective technique lies in mapping the gray levels based on the probability distribution function of an input. Mainly, HE technique flattens and stretches the gray level input image over the dynamic range which results in over enhancement. In smear image, over enhancement reduces the subtle edge features that results in false classification. For such reason, weighted histogram equalization (WHE) technique is adopted to eliminate illumination effect on smear image with certain thresholding conditions. Compared to other HE methods, WHE has better adaptivity and ease of control. Now the quality of an image is enhanced through certain transformation function

T (\cdot)

. Consider the Weighted Histogram Equalization technique (WHE) which is applied over the formulated input and it is given as,

P_{w} (G_{k}) = T {P (G_{k})},

(2)

P_{w} (G_{k}) = {\begin{cases} \hat{P_{w}} (G_{k}), & P (G_{k}) > τ_{1}, \\ 0, & P (G_{k}) > τ_{2}, \end{cases}

(3)

Figure 2 Proposed Smear-net architecture

Full size|PPT slide

τ_{1}

and

τ_{2}

are the upper and lower thresholds which are defined as

τ_{1} \to β^{*} max {P (G_{k})},

(4)

τ_{2} \to G_{k} < 1,

(5)

where

\to

weight function and it takes a value

< 1

and not zero.

Hence, the WHE based on thresholding can be calculated using the formula:

P_{w} (G_{k}) = [(P (G_{k}) - τ_{2}) / (τ_{1} - τ_{2})] * τ_{1} for τ_{1} < τ \leq τ_{2} .

(6)

The image is processed with the above function to handle distinct illumination conditions and considered as one of the inputs for CNN. Edge enhancement technique plays a vital role in image classification. Mainly, the convolutional neural network (CNN) classifies the images based on the edge features. In facial expression recognition, edge enhancement technique enriches the latent features which identifies the subtle expression variations. Hence it is essential to improve and preserve edge features to get better expression classification results. The edge enhancement is performed on the same set of video-sequence using distance map and geometric distance map method which are discussed below.

2.1 Distance Map Method

In distance map method, each pixel in the image is labelled with a distance value, which is nearest to the active neighbourhood pixel. Hence the distance between two prominent edge points is calculated using Euclidean distance metric. The mathematical representation of the Euclidean metric is given in an equation (7).

D_{e u c} (p, q) = \sqrt{\sum_{i = 1}^{n} (p_{i} - q_{i})^{2}} .

(7)

2.2 Geometric Distance Map Method

The geometric distance map describes geometric information of a set of pixels in the 3D space image based on the distance calculated between the member pairs using a chamfer distance metric. The computation is between each foreground voxel to the nearest background voxel. The chamfer distance is the average distance between two points

M

and

N

and is presented in the following equation:

C D (M, N) = 1 / p \sum_{m_{i} \in M} min_{n_{j} \in N} | m_{i} - n_{j} |,

(8)

where

M = {m_{i}}

N = {n_{j}}

and

p

is the number of points in

M

. The distance map and geometric distance map techniques are superimposed to get good edge information as resultant image that assists in detecting expressions better.

2.3 Feature Learning

As mentioned in CNN is composed of convolutional layer, MaxPooling layer with an activation function, and a fully connected layer. In the proposed architecture, after each convolutional layer and before maxpooling, batch normalization (BN) and rectified linear unit (ReLU) activation function are added. The function of batch normalization boosts the network to learn the features very effectively during training and reduces overfitting. Mainly, the effective classification is highly subjective to the image features which are associated with an edge and the depth of the network. In this research, two pre-processing techniques are applied for image sequences to preserve edge features and the connectivity of each edge for better classification. If the input of the 2D convolutional layer is

I (x, y)

, and the corresponding feature map

s (x, y)

will be obtained by convolving the input data with a convolution kernel

w (x, y)

of size

m \times n

s (x, y) = I (x, y)^{*} w (x, y),

(9)

s (x, y) = \sum_{m = 0}^{M - 1} \sum_{n = 0}^{N - 1} I (m, n) \cdot w (x - m, y - n),

(10)

After each convolution operation, the input to the successive hidden layer is changed. Such distributions in all the hidden layers may limit the performance of layer parameters which also reduce the learning rate during training. To address such issue, BN technique is introduced in the proposed architecture to regularize the internal co-variate shift of each convolution layer.

The activations of a given input is normalized before it passes to the next layer in the network. If

x

is considered as the mini-batch of activations, then the normalized input

\hat{x}

can be computed using the following equation,

\hat{x} = (x_{i} - μ_{B}) / \sqrt{σ_{β}^{2} + ϵ} (normalize),

(11)

where

μ_{B}

and

σ_{β}^{2}

are the respective mean and variance over each mini-batch of training images,

β

ϵ

is the error, which set equal to a small positive value. Hence, the activation of convolution layer will have approximately a zero mean and unit variance.

μ_{B} \leftarrow \frac{1}{m} \sum_{i = 1}^{m} x_{i} (Mini-batch Mean),

(12)

σ_{β}^{2} \leftarrow \frac{1}{m} \sum_{i = 1}^{m} (x_{i} - μ_{β})^{2} (Mini batch Variance),

(13)

y_{i} \leftarrow γ \hat{x_{l}} + β \equiv B N_{γ, β} (x_{i}) (Scale and Shift) .

(14)

Hence, the output of the convolution layer is multiplied by a standard deviation (

γ

-gamma) and adds a mean (

β

-beta), which ranges of 0 and 1. The

γ

and

β

are the two parameters learned in each layer which preserves the network ability. The main advantages of using BN as follows, it reduces overfitting, Minimizes the loss, Stabilizes the network by allowing for a wide variety of learning rate and regularization techniques. Now the normalised features obtained using BN are given to the max-pooling layer. The function of maxpooling learns significant value of invariant feature maps. Also, max-pooling layer reduces the dimensional of previous layer output images by a factor of

k_{I (x, y)}

through a non-linear down-sampling technique. Hence maxpooling improves generalisation and reduces overfitting. Here, the sampling size of max-polling is

3 \times 3

. Further, adopted the ReLU (Rectified Linear Unit) non-linear activation function. The use of ReLU increases the non-linearity and avoids the vanishing gradient problem. It is defined as,

R e L U (x) = max (0, x) .

(15)

The neurons of the fully connected layer which connects all the convolutional layer, maxpooling layer, and its activation together. It is a special form of a convolutional layer that acts as a multilayer perceptron. It is given as

F (x) = σ (W^{*} x),

(16)

where

σ

is an identity function,

W

is the weight matrix, and

x

is the input.

Finally, the output from the dual CNN structure are fused and given to Global average pooling layer, which reduces the dimensionality of spatial features globally by averaging all the feature vectors obtained from the previous layer. For further classification, the softmax layer is used. It classifies and predicts the probability of each expression labels based on the obtained feature maps. The softmax layer has an ability to classify a non-linear functions and is given as,

P (x)_{j} = \frac{e^{x_{j}}}{\sum_{i = 1}^{N} e^{x_{i}}} .

(17)

The proposed architecture is composed of five convolutional layers with 32, 64, 128, 256, and 512 filters of size

5 \times 5

, respectively. After each convolutional layer and before maxpooling layer, batch normalization and ReLU activation functions are considered. The main purpose of BN is to reduce overfitting and optimizes the network by increasing the non-linearity. The obtained features from the convolutional layer are further down-sampled using max-pooling layer. The activations of all the previous layers are connected to a fully connected layer^[29]. Further, the fully connected layer of the dual CNN architecture is fused and given to global average pooling (GAP).

S e n s i t i v i t y (S e = T P / (T P + F N)),

(18)

S p e c i f i c i t y (S p = T N / (T N + F P)),

(19)

A c c u r a c y (A c c = (T P + T N) / (T P + F N + T N + F P)) .

(20)

Finally, the Soft-Max layer predicts the expression categories. The use of dual CNN architecture with Pre-processing techniques enhances the edges of each input image and boost the network to learn higher level of feature abstractions. The proposed network is compact in design and has the capability to train faster.

3 Experimental Set Up and Result Discussion

The effectiveness of the proposed integrated model is evaluated on publicly available database. The experiment is carried out in Ubuntu16.04 with Intel^® core

^{T M}

i7, 2.70 GHz processor and 16 GB RAM memory with 8 GB external NVIDIA graphics card. Database Mendeley datasets, MDE-lab datasets, DTU/Herlev and Own datasets collected form hospitals are used, respectively. The DTU/Herlev Pap Smear Database (2005) contains Pap-smear data acquired from images of healthy & cancerous smears coming from the Herlev University Hospital (Denmark). The number of sequences considered from each dataset for the proposed model evaluation. The DTU/Herlev Pap Smear Database details is shown in Table 2.

Table 2 DTU/Herlev Pap smear Database

class	category	cell Type	Number of sample
1	Normal	Superficial squamous epithelial	74
2	Normal	Intermediate squamous epithelial	70
3	Normal	Columnar squamous epithelial	98
4	Abnormal	Mild squamous non-keratinizing dysplasia	182
5	Abnormal	Moderate squamous	146
6	Abnormal	Severe squamous	197
7	Abnormal	Squamous cell carcinoma	150

The proposed model is implemented using KERAS deep learning library. The model is trained initially with each dataset as mentioned above, based on subject-independent and cross-dataset validation method and also evaluated the performance of the proposed model. First, pre-processing techniques are implemented to correct illumination and to improve the edges. Second, employ a data augmentation technique to overcome the issue of limited data, during training. The use of data augmentation technique avoids overfitting and accurately predicts the unseen data. all the datas are combined and trained based on the evaluation method. Also, the datas are fed into the dual CNN layers and the corresponding layer parameters are shown in Table. The parameters are set to 250 epochs with mini-batch size 80 and the initial learning rate as 0.001. Using Stochastic Gradient Decent method, the model parameters are updated with a momentum of 0.95 and the total loss function is optimized. The parameters are fine-tuned at the fully connected layers and the GAP.

Two different experiment protocols are used for analysis: Subject-independent and cross-dataset protocol tasks. In subject-independent method, the images in the dataset have been divided into training set and validation set in a subject independent manner. It is also called as

K

-fold cross validation. Then, the model is iteratively repeated as per the

K

-folds and averaging the recognition accuracy over the

K

-folds. In cross-dataset evaluation one dataset is used for testing and the other for training. Similarly, the model is iteratively trained and evaluated for all the datasets based on

K

-fold cross-validation manner. In this experiment, each dataset follows the 10-fold cross-validation method (i.e., the datasets are split into 10 groups during the CNN training, with 9 groups for training and the remaining 1 group for validation each time and repeat it for 10 times) to evaluate the proposed model. From confusion matrix, the classification accuracy is measured for each expression. From the literature, the evaluation differs for six class and seven class (including neutral) expression classification. Also, the classification accuracy depends on choosing the annotated expression frames for training. The process is carried out with 10-fold cross validation protocol in which the nine folds are used for training and remaining one for validation. In this various folding is used, such as 8-fold and 5-fold cross-validation technique and experimented for 6 class expression labels. The confusion matrix is shown is Table 3. In proposed SMEARNET model, used 10-fold cross-validation and evaluated with benchmark datasets. The number of training samples should be increased to attain better classification. The overall comparisons of the various methods are as shown in the Table 4.

Table 3 Confusion Matrix

	Superficial squamous epithelial	Intermediate squamous epithelial	Columnar squamous epithelial	Mild squamous non-keratinizing dysplasia	Moderate squamous	Severe squamous	Squamous cell carcinoma
Superficial squamous epithelial	98.64	2.78	1.58	0	0	0	0
Intermediate squamous epithelial	2.58	96.87	1.55	0	0	0	0
Columnar squamous epithelial	5.05	2.04	93.48	0.49	3.94	0	0
Mild squamous non-keratinizing dysplasia	0	0	0	95.05	0	4.95	0
Moderate squamous	0	0	2.47	0	92.25	0	6.28
Severe squamous	0	0	0.68	3.48	0	92.52	1.32
Squamous cell carcinoma	2.98	0	0	0	1.75	0	97.27

Table 4 Performance comparison of the proposed method

Classifiers	Accuracy (%)		Error (%)
Classifiers	Normal smear cells	Abnormal smear cells	Normal	Abnormal
Principal component analysis (PCA)	87	81	13	19
Discrete Wavelet Transform Kernel Principal component analysis (DWT-KPCA)	90	85	10	15
Support vector machine (SVM)	92	90	8	10
probabilistic neural network (PNN)	97	93	3	7
Convolutional neural network (CNN)	99.34	98.7	0.7	1.23
Proposed SMEARNET	99.57	99.21	0.41	0.81

Table 5 shows that the overall performance measures of the smear cell segmentation. The parameters of the accuracy, sensitivity and specificity are calculated. The proposed method improved the accuracy, sensitivity and specificity in the classification of smear cells.

Table 5 Overall performance measures

Methods	Accuracy	Sensitivity	Specificity
Semi automatic morphology approach^[8]	92.68	92.23	93.22
Contour property-based approach^[33]	96.8	92.43	93.63
B spline based method^[38]	98.65	94.65	96.71
Geodisc segmentation^[7]	98.85	96.32	95.23
Multi thresholding clustering^[35]	99.39	96.32	95.53
Clustering based Methods^[32]	98.96	97.32	94.32
Fuzzy Based Techniques^[36]	98.81	96.11	94.91
Histogram-based Techniques^[40]	97.12	94.91	93.67
Proposed method	99.57	98.57	97.30

4 Conclusion

In this paper novel Smear-net model is discussed for the analysis of pap-smear cells. The proposed method achieved 95.39 percentage of sensitivity, 90.18 specificity and 96.89 accuracy for

P

-cell segmentation. The proposed method achieved 99.9 percentage of sensitivity, 98.30 in specificity and 98.65 accuracy for

S

-cell segmentation. The classification accuracy of smear-net model is 99.87 for normal cells and for abnormal cells segmentation accuracy of 99.41. The smear-net model provides the classification accuracy of 99.57 for normal cells and for abnormal cells segmentation accuracy of 99.21.

References

Publishing order | Descend order by publishing year | Descend order by cited within

1	Alush A, Greenspan H, Goldberger J. Automated and interactive lesion detection and segmentation in uterine cervix images. IEEE Transactions on Medical Imaging, 2010, 29(2): 488- 501. https://doi.org/10.1109/TMI.2009.2037201 Cited in this article [1]

2	Bamford P, Lovell B. A water immersion algorithm for cytological image segmentation. Proceedings of APRS Image Segmentation Workshop, Sydney, 1996, 75- 79.

3	Barrasso R. Colposcopy as a screening tool for cervical cancer detection: A review. Franco M J. New Developments in Cervical Cancer Screening and Prevention. Oxford: Blackwell Science, Ltd, 1997: 400–405.

4	Bergmeir C, Garcia-Silvente M, Benitez J M. Segmentation of cervical cell nuclei in high-resolution microscopic images: A new algorithm and a web-based software framework. Computer Methods and Programs in Biomedicine, 2012, 107(3): 497- 512. https://doi.org/10.1016/j.cmpb.2011.09.017

Bernstein S J, Sanchez-Ramos L, Ndubisi B. Liquid-based cervical cytologic smear study and conventional Papanicolaou smears: A metaanalysis of prospective studies comparing cytologic diagnosis and sample adequacy. American Journal of Obstetrics and Gynecology, 2001, 185, 308- 317.

https://doi.org/10.1067/mob.2001.116736

Cited in this article [1]

6	Bray F, Ferlay J, Soerjomataram I, et al. Global cancer statistics 2018: Globocan estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA: A Cancer Journal for Clinicians, 2018, 68(6): 394- 424. https://doi.org/10.3322/caac.21492 Cited in this article [1]

7	Chankong T, Theera-Umpon N, Auephanwiriyakul S. Automatic cervical cell segmentation and classification in pap smears. Computer Methods and Programs in Biomedicine, 2014, 113(2): 539- 556. https://doi.org/10.1016/j.cmpb.2013.12.012 Cited in this article [1]

8	Chen Y F, Huang P C, Lin K C, et al. Semi-automatic segmentation and classification of pap smear cells. IEEE Journal of Biomedical and Health Informatics, 2014, 18(1): 94- 108. https://doi.org/10.1109/JBHI.2013.2250984 Cited in this article [2]

9	Somasundaram D. Pixel sensible local band analysis in microscopic chromosome images using CSPA. Cytology and Genetics, 2016, 50, 42- 46. https://doi.org/10.3103/S0095452716010084 Cited in this article [1]

10	Cullins V E, Wright T C, Beattie K J, et al. Cervical cancer prevention using visual screening methods. Reproductive Health Matters, 1999, 7, 134- 143. https://doi.org/10.1016/S0968-8080(99)90014-9

11	Data S D, Koutsky L A, Ratelle S, et al. Human papillomavirus infection and cervical cytology in women screening for cervical cancer in the United States, 2003-2005. Journal of Ann Med, 2008, 148, 493- 500.

12	Database: DTU/Herlev Pap Smear Database: http://labs.fme.aegean.gr/decision/downloads, 11.02.2014.

13	Genctav A, Aksoy A, Onder S. Unsupervised segmentation and classification of cervical cell images. Pattern Recognition, 2012, 45, 4151- 4168. https://doi.org/10.1016/j.patcog.2012.05.006 Cited in this article [1]

14	Gordon S, Zimmerman G, Greenspan H. Image segmentation of uterine cervix images for indexing in PACS. Proceedings 17th IEEE Symp Computer-Based Medical Systems, Bethesda, MD, CBMS, 2004, 298- 303. Cited in this article [1]

15	Gordon S, Zimmerman G, Long R, et al. Content analysis of uterine cervix images: Initial steps towards content based indexing and retrieval of cervigrams. Proc SPIE Conf Medical Imaging, San Diego, CA, 2006, 6144, 1549- 1556.

16	Harandi N M, Sadri S, Moghaddam N A, et al. An automated method for segmentation of epithelial cervical cell images of ThinPrep. Journal of Medical Systems, 2010, 34(6): 1043- 1058. https://doi.org/10.1007/s10916-009-9323-4

17	He K, Gkioxari G, Dollar P, et al. Mask $R$ -CNN. Proceedings of the IEEE International Conference on Computer Vision, 2010, 2961- 2969.

18	Herrero R, Schiffman M H, Bratti C, et al. Design and methods of a population-based natural history study of cervical neoplasia in a rural province of Costa Rica: The Guanacaste Project. Pan American Journal of Public Health, 1997, 1(15): 362- 375.

19	Hofer T P, Katz S J. Health behaviors among women in the United States and Ontario: The effect of use of preventive care. American Journal of Public Health, 1996, 86(12): 1755- 1759. https://doi.org/10.2105/AJPH.86.12.1755 Cited in this article [1]

20	Holowaty P, Miller A B, Rohan T T. Natural history of dysplasia of the uterine cervix. Journal of National Cancer Institute, 1999, 91(3): 252- 258. https://doi.org/10.1093/jnci/91.3.252 Cited in this article [1]

21	LeCun Y, Bengio Y, Hinton G. Deep learning. Nature, 2015, 521(7553): 436. https://doi.org/10.1038/nature14539

Lezoray O, Cardot H. Cooperation of color pixel classification schemes and color watershed: A study for microscopic images. IEEE Transactions on Image Processing A Publication of the IEEE Signal Processing Society, 2002, 11, 783- 789.

https://doi.org/10.1109/TIP.2002.800889

Cited in this article [1]

23	Marinakis Y, Marinaki M, Dounias G. Particle swarm optimization for pap-smear diagnosis. Expert Systems with Applicatonsl, 2008, 35(4): 1645- 1656. https://doi.org/10.1016/j.eswa.2007.08.089 Cited in this article [1]

24	Plissiti M E, Nikou C, Charchanti A. Automated detection of cell nuclei in Pap smear images using morphological reconstruction and clustering. IEEE Trans Inf Technol Biomed, 2011, 15, 233- 241. https://doi.org/10.1109/TITB.2010.2087030 Cited in this article [1]

25	Russakovsky O, Deng J, Su H, et al. Imagenet large scale visual recognition challenge. International Journal of Comput Vision, 2015, 115(3): 211- 252. https://doi.org/10.1007/s11263-015-0816-y Cited in this article [1]

26	Saraiya M, Ahmed F, Krishnan S. Cervical cancer incidence in a prevaccine era in the United States, 1998-2002. Obstetrics and Gynecology, 2007, 109, 360- 370. https://doi.org/10.1097/01.AOG.0000254165.92653.e8

27	Sulaimana S N, Isa N A M, Othmanc N H. Semi-automated pseudo color features extraction technique for cervical cancer's Pap smear images. International Journal of Knowledge-based in Intelligent Engineering Systems, 2011, 15, 131- 143.

28	Zhang L, Lu L, Nogues I, et al. Deep-pap: Deep convolutional networks for cervical cell classification. IEEE Journal of Biomedical and Health Informatics, 2017, 21(6): 1633- 1643. https://doi.org/10.1109/JBHI.2017.2705583 Cited in this article [1]

29	Rajan S, Chenniappan P, Devaraj S, et al. Novel deep learning model for facial expression recognition based on maximum boosted CNN and LSTM. IET Image Processing, 2020, 14, 1373- 1381. https://doi.org/10.1049/iet-ipr.2019.1188 Cited in this article [2]

30	Kumaresan N, Somasundaram D. Review of pap smears cell segmentation and classification techniques for cervical cancer analysis. Studies on Ethno Medicine, 2018, 12(2): 96- 105. https://doi.org/10.31901/24566772.2018/12.02.544

31	Devaraj S, Vijaykumar V R, Soundrarajan G R, et al. Leaf biometrics based karyotyping of G-band chromosomes. International Journal of Human Genetics, 2013, 13(3): 131- 138. https://doi.org/10.1080/09723757.2013.11886208

32	Cheng J, Rajapakse J C. Segmentation of clustered nuclei with shape markers and marking function. IEEE Transactions on Biomedical Engineering, 2009, 56(3): 741- 748. https://doi.org/10.1109/TBME.2008.2008635 Cited in this article [2]

33	Rinneburger M, Carolus H, Iuga A I, et al. Automated localization and segmentation of cervical lymph nodes on contrast-enhanced CT using a 3D foveal fully convolutional neural network. European Radiology Experimental, 2023, 7(1) https://doi.org/10.1186/s41747-023-00360-x Cited in this article [2]

34	Somasundaram D, Gnanasaravanan S, Madian N. Automatic segmentation of nuclei from pap smear cell images: A step toward cervical cancer screening. International Journal of Imaging Systems and Technology, 2020, 30, 1209- 1219. https://doi.org/10.1002/ima.22444

35	Somasundaram D, Kumaresan N, Subramanian V, et al. Structural similarity and probabilistic neural network based human g-band chromosomes classification. International Journal of Human Genetics, 2018, 18(3): 228- 237. Cited in this article [1]

36	William W, Ware A, Basaza-Ejiri A H, et al. Cervical cancer classification from Pap-smears using an enhanced fuzzy C-means algorithm. Informatics in Medicine Unlocked, 2019, 14, 23- 33. https://doi.org/10.1016/j.imu.2019.02.001 Cited in this article [2]

37	Chankong T, Theera-Umpon N, Auephanwiriyakul S. Automatic cervical cell segmentation and classification in Pap smears. Comput Methods Programs in Biomedicine, 2014, 113(2): 539- 556. https://doi.org/10.1016/j.cmpb.2013.12.012 Cited in this article [1]

38	Bai X Z, Sun C M, Zhou F G. Touching cells splitting by using concave points and ellipse fitting. Proceedings of the 2008 Digital Image Computing: Techniques and Applications, 2008, 271- 278. Cited in this article [1]

39	Madian N, Jayanthi K B, Somasundaram D, et al. Identifying centromere position of human chromosome images using contour and shape based analysis. Measurement, 2019, 144, 243- 259. https://doi.org/10.1016/j.measurement.2019.05.029 Cited in this article [1]

40	Chen B, Mete N, Kockara S. Parameter free multilevel fuzzy c mean clustering for unsupervised detection in histological images. Society for Design and Process Science, USA, Systems Conference, 2010. Cited in this article [1]

PDF(724 KB)

542

Accesses

Citation

Detail

Sections

Recommended

Abstract
Key words
Cite this article
1 Introduction
2 Proposed Method
Figure 1 Overall pap-smear cell classification
Table 1 Smearnet Architecture
Figure 2 Proposed Smear-net architecture
2.1 Distance Map Method
2.2 Geometric Distance Map Method
2.3 Feature Learning
3 Experimental Set Up and Result Discussion
Table 2 DTU/Herlev Pap smear Database
Table 3 Confusion Matrix
Table 4 Performance comparison of the proposed method
Table 5 Overall performance measures
4 Conclusion
References

Received	Accepted	Published
2023-05-31	2023-11-16	2024-02-25
Issue Date
2024-02-28

Please choose a citation manager

Content to export

Abstract

Key words

Cite this article

1 Introduction

2 Proposed Method

Figure 1 Overall pap-smear cell classification

Table 1 Smearnet Architecture

Figure 2 Proposed Smear-net architecture

2.1 Distance Map Method

2.2 Geometric Distance Map Method

2.3 Feature Learning

3 Experimental Set Up and Result Discussion

Table 2 DTU/Herlev Pap smear Database

Table 3 Confusion Matrix

Table 4 Performance comparison of the proposed method

Table 5 Overall performance measures

4 Conclusion

{{custom_sec.title}}

{{custom_sec.title}}

References

{{custom_fnGroup.title_en}}

Footnotes

Share

模态框（Modal）标题

Please choose a citation manager

Content to export

Abstract

Key words

Cite this article

1 Introduction

2 Proposed Method

Figure 1 Overall pap-smear cell classification

Table 1 Smearnet Architecture

Figure 2 Proposed Smear-net architecture

2.1 Distance Map Method

2.2 Geometric Distance Map Method

2.3 Feature Learning

3 Experimental Set Up and Result Discussion

Table 2 DTU/Herlev Pap smear Database

Table 3 Confusion Matrix

Table 4 Performance comparison of the proposed method

Table 5 Overall performance measures

4 Conclusion

{{custom_sec.title}}

{{custom_sec.title}}

References

{{custom_fnGroup.title_en}}

Footnotes