Detection and classification of cervical cancer images using CEENET deep learning approach

Abstract

Earlier detection of cervical cancer in women can save their lives before a chronic development. The accurate detection in cancer tissues of cervix in the human body is very important. In this article, cervical images were classified into either affected or healthy images using deep learning architecture. The proposed approach was designed with the modules of Edge detector, complex wavelet transform, feature derivation and Convolutional Neural Networks (CNN) architecture with segmentation. The edge pixels in the source cervical image were detected using Kirsch’s edge detector, the Complex Wavelet Transform (CWT) was there used to decompose the edge detected cervical images into number of sub bands. Local Derivative Pattern (LDP) and statistical features were computed from the decomposed sub bands and feature map was constructed using the computed features. The featured map along with the source cervical image was fed into the Cervical Ensemble Network (CEENET) model for classifying of cervical images into the classes healthy or cancer (affected).

Keywords

Cervix deep learning CNN cervical image cancer

1 Introduction

Cervical cancer and breast cancer are the harmful diseases in women patients around the world. When compared to the breast cancer, cervical cancer is the life killing disease which is found in women patients. Some tissues in the cervix region of the women are abruptly changed and produce for the cancer cells. In early stage of this cancer, there are no symptoms found and so the women patient does not feel any pain in their abdomen [1 –4]. In the advanced stage of this cancer, the women patients feel sudden pain in their abdomen followed by bleeding from the cervix. If it is not timely treated, sudden death was occur. Due to transformation from moderate stage to advance of this cancer, the cells in the cervix region spreads rapidly the other parts of the human body such as liver and bladder. Therefore, an earlier detection of this cancer will be an important aspect to save the life of the women [5 –7]. There are two conventional methods available for the identification of cervical cancer namely Cervigram and Pap smear cell test. The Cervigram is a method which scans the cervix region of the women; images scanned are used to detect the cancer regions. Pap smear cell test method, a small portion of the cells in cervix regions that are used to detect the cancer by examining the nucleus in these cells. Arora et al. [16] used Pap smear cells to identify the cervical cancer by machine learning algorithms. The cancer segmentation accuracy level was found to be 92% which is not suitable for further diagnosis and treatment process [9 –11]. Thus, Cervigram method is used in this article to identify the cancer cells. Figure 1(a) shows the cervical image with cancer cells and Fig. 1(b) shows the healthy cervical image.

Fig. 1

(a) Cervical image with cancer cells (b) Healthy cervical image.

Deep learning algorithms have many significant advantages runs on computer vision based pattern identification and health monitoring based automation applications. In computer vision based applications, medical image processing is used for analyzing and segmenting different patterns for different modality of images such as Computer Tomography (CT) and Magnetic Resonance Imaging and Ultra Sound (US) nowadays. Due to these advantages of deep learning algorithms, this article uses a novel deep learning architecture to classify the cervical images into either normal or abnormal.

Section 2 of this article states the survey about conventional methods for cervical cancer detection, section 3 states materials and methods, section 4 elaborates the experimental results with discussions and section 5 concludes this article.

2 Literature survey

Venkatesan Chandran et al. [12] structured a novel methodology for the identification of cervical cancer in colonoscopy images. The authors constructed Colposcopy Ensemble Network (CYENET) architecture for segmenting the cancer pixels present in images. The Visual Geometry Group (VGG)-19 architecture was used by the author to check the effectiveness of the cancer detection process in addition. This proposed method provided sensitivity, specificity, and kappa scores of 92.4%, 96.2%, and 88% respectively for the cervical colonoscopy images in open access dataset. Tripathi et al. [15] used ResNet-152 CNN architecture for the detection and segmentation of cancer regions in cervical images. The authors obtained approximately 95% of classification accuracy using their proposed architecture model for the detection of cancer pixels in cervical images. The developed method was applied on SIPAKMED pap-smear image dataset and validated by the k-fold validation approach. Arora et al. [16] developed cervical cancer detection system using soft computing approach. The authors used polynomial based SVM classification structure for identifying and classifying the cancer affected cervical images from the healthy one. The Gaussian Fitting Energy features were derived from the source cervical image and these features were fed into SVM for the classification process. The cancer pixels in the abnormal cervical image were segmented using Active contour models and authors obtained 96.1% of segmentation accuracy using their developed approach. Chen et al. (2021) developed CytoBrain deep learning model for screening the cancer pixels in the cervical images.

The developed CytoBrain model improved the detection accuracy of the cervical images affected by cancer through the Softmax process present in the CytoBrain model. This work was tested on large number of cervical images for the verification of the developed cervical cancer detection model.

Ghoneim et al. [13] combined CNN architecture with extreme learning machine classification algorithm for the detection and classification of cervical images into sub classes. Authors computed the non-regressive feature set from the source cervical image. these non-regressive feature maps was fed into integrated classification system to classify the image into either cancer or healthy case. Authors validated their developed method on open access dataset images and obtained 96.1% of sensitivity and 97.8% of specificity. Dongyao Jia et al. [14] integrated CNN classification architecture and the machine learning structure Support Vector Machine (SVM) to classify the source cervical image. The CNN architecture was developed and modeled to compute the feature maps from the cervical images and the developed feature path computed the strong feature maps using the prior knowledge. The SVM classification structure performed the dimensionality reduction for improving the classification accuracy of the cervical cancer detection system.

Lu et al. (2020) used Ensemble Machine Learning (EML) algorithm for identifying the cancer affected cervical images. The learning rate of the EML algorithm was based on the computational feature parameters and the classification rate was improved by up scaling the learning rate of the developed EML algorithm in this work. Authors obtained 95.9% of classification rate with mean cancer segmentation accuracy of 96.3% . Saini et al. (2020) constructed ColpoNet Deep learning classification architecture for determining the region of pixels belonging to cancer in cervical images. The constructed ColpoNet was tested on the cervical images with different dataset for analyzing the effect of the developed deep learning structure. Authors modified the internal layers of the developed ColpoNet for improving the cervical cancer classification rate. This method was tested with both manual and computed assisted validation methods. Adem et al. (2019) devised a deep learning network model which consisted both encoder and decoder module for locating the cancer pixels in cervical images. The developed auto encoder model detected and segmented the pixels belonging to the region of cancer in cervical images. The authors validated their developed method on open access dataset images and obtained 96.7% of sensitivity and 97.2% of specificity. Karthiga et al. (2018) developed a machine learning computer simulation model which consist of image registration method along with Adaptive Neuro Fuzzy Inference System (ANFIS) classifier. The authors used Fast Fourier Transform (FFT) for registering the inclined cervical images for image registration process and trinary and texture features were computed from this cervical registered image. These non-linear features were used to locate the cancer affected cervical images through the training of ANFIS classifier.

Based on the conventional methods stated in this section for cervical cancer detection in cervical images, the novelty of the paper is stated as follows.

Instead of processing all pixels in cervical images for detecting cancer pixel, the proposed model stated in this paper process edge pixels only which are detected through the Kirsch’s edge detector. This significantly reduces the cancer pixels detection time.

The new deep learning model CEENET is developed in this paper to classify the cervical images into non-cancer and cancer affected images. The contributions of this proposed model are stated as follows.

Providing an edge detection algorithm for detecting the edge pixels in the cervical images.

Using CWT for decomposition and texture features are computed from the CWT decomposed bands for training the classifier model.

Proposing a novel deep learning model CEENET for the classification of cervical images using the feature maps.

3 Materials and methods

3.1 Materials

The cervical images used in this work are accessed from the Guanacaste dataset. This dataset was established by National Cancer Institute in the year of 1997 for improving the research activities in cervical cancer detection [11]. The cervical images used in this paper are accessed from the link [17] on 12-3-2020. This dataset is consisting of the cervical images belonging to normal and the cervical images belonging to cancer or abnormal. 750 numbers of cervical images from 1024 women patients are obtained and they are categorized into normal (400 number of cervical images) and cancer (350 number of cervical image). 40% of these cervical images are trained and the remaining 60% of cervical images are tested by the proposed method. Therefore, the training dataset consists 160 counts of normal cervical images and 240 counts of cancer images. The testing dataset contains 140 counts of normal cervical images and 210 counts of cancer cervical images. All the cervical images in both training and testing are having 256*256 width and height with 8 bit pixel resolution.

This paper also uses Multi State Colposcopy Image (MSCI) dataset [23, 24] for analyzing the proposed method. This dataset was constructed by University of Science and Technology of China (UST). All the cervical images are acquired by the electronic colposcopy-TR6000 G scanning equipment. The cervical images used in this paper are accessed from the link [27] on 17-4-2020. This dataset was constructed by screening 679 number of women patients in UST university. This dataset consist of 282 numbers of healthy cervical images and 72 numbers of cancer affected cervical images. All the cervical images are having the resolution of 1440*1080 pixels. The cervical images in this dataset are split into 40% of training and 60% of testing mode. Therefore, the training mode of the healthy and cancer affected cervical images consist of 113 images and 29 images respectively. The testing mode of the healthy and cancer affected cervical images consist of 169 images and 243 images respectively.

3.2 Methods

In this article, cervical images are classified into either cancer or healthy images using deep learning architecture. The proposed approach is designed with the modules of Edge detector, complex wavelet transform, feature derivation and CNN architecture with segmentation. The edge pixels in the source cervical image are detected by Kirsch’s edge detector and then CWT is used to decompose the edge detected cervical image into number of sub bands. Local Derivative Pattern (LDP) and statistical features are computed from the decomposed sub bands the feature map is constructed using the computed features. This feature map along with the source cervical image is fed into the CEENET model for the classification of cervical images into various classes healthy or cancer. The entire flow of the proposed model is depicted in Fig. 2.

Fig. 2

Proposed schematic approach for cervical cancer detection.

3.3 Edge detection

Edges represent the abrupt variation of pixel with respect to nearby or surrounding pixels in image. The intensity level of cancer pixels are different from the intensity level of the non-cancer pixels in an image. Hence, the detection of edge pixels in cervical image is important for the detection of cancer region. Sobel edge detector was mostly used in many conventional edge detection processes. This Sobel operator detects the edge pixel in an image using derivative approximation method. The non linearity in derivative approximation method reduces the accuracy level of the detected edge pixels in the image. In order to improve the edge detection process, Kirsch’s edge detector is used in this article to detect the edge pixels with respect to all orientation in an image. This method uses eight different compass filters to find the edge pixels in the cervical image. These eight compass filters are derived from the base compass Convolution filter as mentioned in Shekar et al. [9].

The compass filter is designed with individual kernel and the eight kernels of all compass filters that (Venmathi et al. [10]) are given in Table 1. The size of the kernel matrix of each compass filter is 3*3 and this kernel matrix is convolved with source cervical image in order to produce the response images as illustrated in Fig. 3.

Table 1
Kirsch’s matrix for edge detection process

Kernel name Kernel matrix

K1 $\begin{matrix} + 5 & + 5 & + 5 \\ - 3 & 0 & - 3 \\ - 3 & - 3 & - 3 \end{matrix}$

K2 $\begin{matrix} + 5 & + 5 & - 3 \\ + 5 & 0 & - 3 \\ - 3 & - 3 & - 3 \end{matrix}$

K3 $\begin{matrix} + 5 & - 3 & - 3 \\ + 5 & 0 & + 5 \\ + 5 & - 3 & - 3 \end{matrix}$

K4 $\begin{matrix} - 3 & - 3 & - 3 \\ - 3 & 0 & - 3 \\ + 5 & + 5 & - 3 \end{matrix}$

K5 $\begin{matrix} - 3 & - 3 & + 5 \\ - 3 & 0 & + 5 \\ - 3 & - 3 & + 5 \end{matrix}$

K6 $\begin{matrix} - 3 & + 5 & + 5 \\ - 3 & 0 & + 5 \\ - 3 & - 3 & - 3 \end{matrix}$

K7 $\begin{matrix} - 3 & - 3 & - 3 \\ - 3 & 0 & - 3 \\ + 5 & + 5 & + 5 \end{matrix}$

K8 $\begin{matrix} - 3 & - 3 & + 5 \\ - 3 & 0 & + 5 \\ - 3 & + 5 & + 5 \end{matrix}$

Kernel name	Kernel matrix
K1	$\begin{matrix} + 5 & + 5 & + 5 \\ - 3 & 0 & - 3 \\ - 3 & - 3 & - 3 \end{matrix}$
K2	$\begin{matrix} + 5 & + 5 & - 3 \\ + 5 & 0 & - 3 \\ - 3 & - 3 & - 3 \end{matrix}$
K3	$\begin{matrix} + 5 & - 3 & - 3 \\ + 5 & 0 & + 5 \\ + 5 & - 3 & - 3 \end{matrix}$
K4	$\begin{matrix} - 3 & - 3 & - 3 \\ - 3 & 0 & - 3 \\ + 5 & + 5 & - 3 \end{matrix}$
K5	$\begin{matrix} - 3 & - 3 & + 5 \\ - 3 & 0 & + 5 \\ - 3 & - 3 & + 5 \end{matrix}$
K6	$\begin{matrix} - 3 & + 5 & + 5 \\ - 3 & 0 & + 5 \\ - 3 & - 3 & - 3 \end{matrix}$
K7	$\begin{matrix} - 3 & - 3 & - 3 \\ - 3 & 0 & - 3 \\ + 5 & + 5 & + 5 \end{matrix}$
K8	$\begin{matrix} - 3 & - 3 & + 5 \\ - 3 & 0 & + 5 \\ - 3 & + 5 & + 5 \end{matrix}$

Fig. 3

Response images of Kirsch’s Edge Detector.

The final kirsch’s edge detected image is obtained by finding the maximum intensity of pixel in each response image using the following equation. $\begin{matrix} E (s, t) \\ = Max {| E_{1} (s, t) |, | E_{2} (s, t) | \dots \dots \dots | E_{8} (s, t) |} \end{matrix}$

Where, |E_i (s, t) | = C (i, j) * K_i; i = 1,2. . . .8.

Whereas, C (i, j) is the source cervical image and K_i is the kernel matrix of compass filter and E_i (s, t) is the convolved response image.

3.4 Complex wavelet transform (CWT)

The spatial property of each pixel in edge detected image was improved using transformation process. There are many transformations such as Discrete Wavelet Transform (DWT) and Gabor and Contourlet transforms which are used for the transformation of pixels. Though this method effectively transforms the pixel properties, the redundancy level during pixel transformation is high, but affects the further classification process. In order to reduce the level of redundancy during transformation of pixels in edge detected image, CWT is used in this article. CWT is the type of wavelet transform which is mainly used for low directional selectivity and shift variance. In this article, Dual Tree Complex Wavelet Transform (DTCWT) is used for pixel transformation. The structure of the DTCWT is depicted in Fig. 4. In Fig. 4, the edge detected image is represented as ‘x’ the low and high pass filters are represented as h (n) and g (n) respectively. Three level decomposition structures are used down sampling factor 2 at each stage.

Fig. 4

DTCWT structure for decomposition of edge detected image.

DTCWT can be designed using complex wavelet function and complex scaling function. The design of the DTCWT is given in the following equations. $ϑ (t) = ϑ_{h} (t) + j ϑ_{g} (t)$ $\emptyset (t) = \emptyset_{h} (t) + j \emptyset_{g} (t)$

Where as, ϑ_h (t) and ϑ_g (t) are the real wave function and imaginary wave function respectively and ∅_h (t) and ∅_g (t) are the real scale function and imaginary scale function respectively.

In this article, eight decomposed sub band images are obtained from the edge detected cervical image.

3.5 Feature extractions

The textural contents of each decomposed sub band images are used to differentiate the normal cervical image from the cancer cervical image. These textural contents of each sub band image are computed using feature extraction method. In this article, LDP and statistical features are computed from each decomposed sub band image for further classification process. These features are explained in the following sections.

3.5.1 LDP features

The LDP feature is originally derived from the basics of Local Binary Pattern (LBP), which works based on the intensity variation of each pixel with respect to the surrounding pixels. In LBP computation, the pixel is replaced with gradient value which generates the non-stable and non-linearity in further computational process. This reduces the classification process which is based on the feature extraction. In order to overcome such limitation in LBP, the LDP feature (Abuobayd et al. [11]) extraction process is used in this article to compute the distinct features from each decomposed sub band image.

The LDP of each pixel in decomposed sub band image is computed with respect to four directional using the following equation. ${LDP}_{θ} (P_{c}) = \sum_{i = 1}^{P} 2^{i - 1} * S [(D_{θ}^{k - 1} (P_{c}), D_{θ}^{k - 1} (P_{n})]$

Where as, P_c is the center pixel in 3*3 window of each decomposed sub band image and P is the number of surrounding pixels in 3*3 window with neighboring pixel P_n. The orientation of each pixel in sub band image is depicted by θ and it has the values of 0⁰, 45⁰, 90⁰, 135⁰. The scaling function is represented by S and it is computed using the following equation. $S (x, y) = {\begin{matrix} 1; if x * y \geq 0 \\ 0; if else \end{matrix}$

Figure 5(a–d) shows the computed LDP feature image from the decomposed sub band image.

Fig. 5

(a) LDP image at 0⁰ (b) LDP at 45⁰ (c) LDP at 90⁰ (d) LDP at 135⁰.

3.5.2 Statistical features

The statistical features are computed from each decomposed sub band image which highlights the statistical variance of each pixel. These statistical variances are used to differentiate the normal cervical image from the cancer cervical image. These statistical features are computed by the following equations. $Mean (\bar{M}) = \frac{\sum_{i = 1}^{P} \sum_{j = 1}^{Q} M (i, j)}{N}$

Where, M (i, j) is the sub band matrix and N is the total number of elements in sub band matrix. $Variance (σ^{2}) = \frac{\sum_{i = 1}^{P} \sum_{j = 1}^{Q} [M (i, j) - \bar{M}]^{2}}{N}$ $Skewness = \frac{1}{N} \sum_{i = 1}^{P} \sum_{j = 1}^{Q} {[\frac{M (i, j) - (\bar{M})}{σ}]}^{3}$ $Kurtosis = \frac{1}{N} \sum_{i = 1}^{P} \sum_{j = 1}^{Q} {[\frac{M (i, j) - (\bar{M})}{σ}]}^{4}$

3.6 Classifications

The computed features are classified as conventional machine learning method or proposed deep learning architecture. Mostly, many researchers used machine learning methods Adaptive Neuro Fuzzy Inference System (ANFIS) and Evolutionary approaches for the classification of cervical images. The classification accuracy rates of these conventional machines learning based methods are not optimum. There is a need for developing a methodology which provides optimum classification accuracy rate. Therefore, the study develops a novel architecture for the deep learning structure for the classification of cervical images.

Figure 6 shows the proposed flow chart of the cervical cancer detection using CNN model. The cervical images from dataset (both normal and abnormal) that are data augmented (Shift left and shift right functions are used as the data augmentation methods). They are split into training dataset and testing dataset. The cervical images from the training dataset along with the external feature maps (derived LDP and statistical features) are trained by the proposed CNN model which produces the training patterns. The source cervical image from the testing dataset is classified using proposed CNN model along with the trained patterns.

Fig. 6

Proposed flow chart of the cervical cancer detection using CNN model.

Table 2 shows the specifications of the proposed CEENET model with respect to number of Convolutional layers, pooling layers, and numbers of filters, stride and number of neurons. The negative responses are removed using ReLU layer. This is activation layer which has the piecewise linear function as described in the following equation.

Table 2

Specifications of the proposed CEENET model

Layer type	Filter kernel size	stride	Number of filters	Number of neurons
Convolution_1	7×7	2×2	32	–
Maxpool_1	3×3	2×2	–	–
Convolution_2	5×5	1×1	64	–
Maxpool_2	3×3	2×2	–	–
Convolution_3	5×5	1×1	128	–
Maxpool_3	3×3	2×2	–	–
Convolution_4	3×3	1×1	256	–
Maxpool_4	3×3	2×2	–	–
Convolution_5	3×3	1×1	256	–
Maxpool_5	3×3	2×2	–	–
Maxpool_6	3×3	2×2	512	–
Fully Connected Neural Network (FCNN)-1	–	–	–	1024
Fully Connected Neural Network (FCNN)-2	–	–	–	2

$R (x) = \max (0, x)$

Whereas, x is the input to the ReLU layer.

The responses from both FCNN1 and FCNN2 are optimized using SOFTMAX layer in the proposed CEENET model. The softmax function of the SOFTMAX layer is designed using the following equation. $S (z) = \frac{e^{z_{i}}}{e^{z_{g}}}$

Whereas, z_i and z_g are the FCNN-1 and FCNN-2 outputs respectively and S(x) is the softmax response of both FCNN-1 and FCNN-2 layers.

The morphological functioning (dilation followed by erosion) used in this article which is applied on the classified of abnormal cervical image to segment the pixels belonging to the cancer category.

Figure 7 shown the structure of CEENET model which is proposed in this article for the classification of cervical images. This CEENET model is derived from the conventional Model CYENET (Venkatesan Chandran et al. [12]). The conventional CYENET model required high number of Convolutional layers and pooling layers, which makes the proposed design more complex and time consuming process. These limitations are eliminated by modifying the structure of CYENET model into CEENET model. In Fig. 7, five numbers of Convolutional layers, six numbers of pooling layers and two numbers of FCNN layers are used. The first Convolutional layer (Convolution_1) convolves (multiplication) the computed features with the kernel of this layer. The size of the response of this layer is high due to the process of Convolution. This response may contain negative elements which can be eliminated by passing this response through Linear Rectification Layer (ReLU_1 layer). The size of this response from this layer is reduced by pooling layer (MaxPool_1). The feature maps from the MaxPool_1 are now passed through Convolution_2, ReLU_2 and MaxPool_2 in block 2 represented is Fig. 7. The feature maps from the MaxPool_2 are now passed through Convolution_3, ReLU_3 and MaxPool_3 in block 3 of Fig. 7. Now, the response from the MaxPool_3 is passed simultaneously through block 4 and block 5 as shown in Fig. 7. The feature maps which are generated by MaxPool_4 and MaxPool_5 are integrated and passed through MaxPool_6 to reduce the size of this feature maps. Finally, this feature map is passed through the two FCNN layers to produce the final classification results.

Fig. 7

Proposed CEENET model.

Further, morphological segmentation method segments the cancer pixels in the classified abnormal cervical images. Figure 8(a) shown the source cervical images and Fig. 8 (b) shows the cancer pixels segmented images by proposed method stated in this paper.

Fig. 8

(a) Cervical images (b) Cancer pixels segmented image by proposed method.

4 Results and discussions

Guanacaste and MSCI datasets are used in this work to analyze the performance efficacy of the proposed system. From the Guanacaste dataset, 240 cervical images belonging to normal category and 210 cervical images belonging to cancer category are tested using the proposed method. From the MSCI dataset, 169 cervical images belonging to normal category and 43 cervical images belonging to cancer category are tested using the proposed method stated in this article.

The performance of this proposed method is estimated using the parameter Normal Index Rate (NIR) and Cancer Index Rate (CIR). The NIR is defined as the ratio between the counts of the normal cervical images correctly detected to the total count of the normal cervical images. The CIR is defined as the ratio between the counts of the cancer cervical images correctly detected to the total count of the cancer cervical images. Both NIR and CIR are measured in percentage and the performance of the cervical cancer detection is high if both measured NIR and CIR values are high. In this article, 237 normal cervical images are correctly detected by the proposed system and hence the proposed system achieves 98.7% of NIR. Also, 207 cancer cervical images are correctly detected by the proposed system and hence the proposed system achieves 98.5% of NIR. Therefore, 98.6% of Mean Index Rate is obtained by the proposed method on Guanacaste dataset. In this article, 167 normal cervical images are correctly detected by the proposed system and hence the proposed system achieves 98.8% of NIR for MSCI dataset images. 42 cancer cervical images are correctly detected by the proposed system and the proposed system achieves 97.6% of NIR for MSCI dataset images. Therefore, 98.2% of Mean Index Rate is obtained by the proposed method on MSCI dataset.

In addition, the proposed cervical cancer detection methodology is experimentally analyzed with respect to the metrics stated in the Equations. $Sensitivity (Sey) = \frac{TP}{TP + FN}$ $Specificity (Spy) = \frac{TN}{TN + FP}$ $\begin{matrix} TumorDetectionAccuracy (TDA) \\ = \frac{TP + TN}{TP + TN + FP + FN} \end{matrix}$ $PositivePredictiveRate (PPR) = \frac{TP}{TP + FP}$ $NeagativePredictiveRate (NPR) = \frac{TN}{TN + FN}$

Where as, the actual detected cancer and non-cancer pixels in cervical image are denoted by TP and TN respectively. The falsely detected cancer and non-cancer pixels in cervical image are denoted by FP and FN respectively.

The sensitivity computes percentage of pixels belonging to cancer category with respect to the positive results. The specificity computes percentage of pixels belonging to normal or non-cancer category with respect to the negative results. TDA defines the percentage of pixels belonging to correctly identified cancer region in abnormal image. The parameters PPR and NPR determine the percentage of correctly detected tumor and non-tumor pixels in classified cervical image, respectively. These parameters are estimated by the proposed cervical cancer detection method using manually cancer segmented cervical images.

Table 3 is the experimental evaluation of the cervical cancer detection system using the methodology stated in this article on Guanacaste dataset. This proposed method is tested on the cancer cervical images (10 numbers of cervical images) and obtained 98.3% of sensitivity, 98.3% of specificity, 99.1% of TDA, 98.1% of PPR and 98.2% of NPR. The similar experimental results are also obtained by testing the proposed method on the entire number of cervical images (350 abnormal cervical images) in Guanacaste dataset in this paper.

Table 3
Experimental evaluation of the cervical cancer detection system on Guanacaste dataset

Cervical images Sey Spy TDA PPR NPR

1 98.5 97.9 99.2 98.3 98.3

2 98.3 98.2 99.6 98.1 97.7

3 98.2 98.9 99.3 98.9 97.3

4 99.1 98.3 99.1 97.5 98.7

5 98.5 98.1 98.9 97.9 98.8

6 97.9 97.9 98.7 97.5 97.7

7 98.3 98.4 99.3 98.2 98.7

8 97.2 98.2 99.2 98.7 98.3

9 98.7 98.8 99.1 98.3 98.2

10 98.3 98.3 98.9 97.9 98.7

Mean 98.3 98.3 99.1 98.1 98.2

Cervical images	Sey	Spy	TDA	PPR	NPR
1	98.5	97.9	99.2	98.3	98.3
2	98.3	98.2	99.6	98.1	97.7
3	98.2	98.9	99.3	98.9	97.3
4	99.1	98.3	99.1	97.5	98.7
5	98.5	98.1	98.9	97.9	98.8
6	97.9	97.9	98.7	97.5	97.7
7	98.3	98.4	99.3	98.2	98.7
8	97.2	98.2	99.2	98.7	98.3
9	98.7	98.8	99.1	98.3	98.2
10	98.3	98.3	98.9	97.9	98.7
Mean	98.3	98.3	99.1	98.1	98.2

Table 4 shows the experimental results of the proposed method on the cervical images from MSCI dataset in this article. This proposed method is tested on the cancer cervical images (10 numbers of cervical images) and obtained 99.06% of sensitivity, 99.34% of specificity, 99.32% of TDA, 98.98% of PPR and 98.96% of NPR.

Table 4

Experimental evaluation of the cervical cancer detection system on MSCI dataset

Cervical images	Sey (%)	Spy (%)	TDA (%)	PPR (%)	NPR (%)
1	99.2	99.3	99.3	99.3	98.3
2	99.1	99.1	99.6	99.1	99.3
3	98.9	98.9	99.5	98.9	98.6
4	99.5	99.7	99.3	99.3	98.7
5	99.1	99.5	99.1	98.5	98.8
6	98.5	99.4	99.5	98.7	99.4
7	98.9	99.4	99.3	99.2	98.7
8	98.7	99.3	99.2	99.3	99.1
9	99.3	99.7	99.1	98.3	99.4
10	99.4	99.1	99.3	99.2	99.3
Mean	99.06	99.34	99.32	98.98	98.96

The similar experimental results are also obtained by testing the proposed method on the entire number

of cervical images (72 abnormal cervical images) in MSCI dataset in this paper.

In this article, the CEENET based cervical cancer detection system is experimentally evaluated by various transforms DTCWT (in this work), DWT (Pushpalatha et al. 2018) and Contourlet (Soumya et al. 2016), as illustrated in Table 5 on both Guanacaste and MSCI datasets respectively. From the analysis of implementation of various transforms on cervical cancer detection system, DTCWT achieved better cancer region segmentation results than the other transforms in this article. The cervical images in MSCI dataset are having low pixel intensity then the cervical images in Guanacaste dataset. From this Table 5, it is observed that the DTCWT for Guanacaste dataset obtains high performance experimental results when compared with the DTCWT for MSCI dataset.

Table 5

Experimental evaluation of the cervical cancer detection system using various transforms

Datasets	Transforms	Sey (%)	Spy (%)	TDA (%)	PPR (%)	NPR (%)
Guanacaste	DTCWT (in this article)	98.3	98.3	99.1	98.1	98.2
DWT	91.8	90.4	92.9	91.8	90.6
Contourlet	92.3	91.6	91.7	91.6	91.7
MSCI	DTCWT	99.06	99.34	99.32	98.98	98.96
DWT	92.1	92.6	93.2	92.6	93.2
Contourlet	91.9	91.7	91.9	92.5	92.6

In this article, the cervical cancer detection system is experimentally evaluated by deep learning models CEENET (in this article) and CYENET (Venkatesan Chandran et al. 2021), as illustrated in Table 6 on both Guanacaste and MSCI datasets respectively. From the analysis of implementation of various deep learning models on cervical cancer detection system, the proposed CEENET model stated in this article achieved better cancer region segmentation results than the other transforms in this article.

Table 6

Experimental evaluation of the cervical cancer detection system using CEENET and CYENET models

Datasets	Transforms	Sey (%)	Spy (%)	TDA (%)	PPR (%)	NPR (%)
Guanacaste	CEENET (in this article)	98.3	98.3	99.1	98.1	98.2
CYENET (Venkatesan Chandran et al. 2021).	95.9	96.2	96.5	96.2	96.7
MSCI	CEENET (in this article)	99.06	99.34	99.32	98.98	98.96
CYENET (Venkatesan Chandran et al. 2021).	96.2	97.1	97.4	97.9	97.3

Table 7 shows the comparative analysis of cervical cancer detection with other similar methods on both Guanacaste and MSCI datasets. Venkatesan Chandran et al. (2021) obtained 92.4% of Sey, 96.2% of Spy. Tripathi et al. (2021) obtained 95.3% of Sey, 96.9% of Spy and 96.8% of TDA. Dongyao Jia et al. (2020) obtained 93.9% of Sey, 97.1% of Spy and 96.9% of TDA on Guanacaste dataset.

Table 7

Comparative analysis of cervical cancer detection with other similar methods on both Guanacaste and MSCI datasets

	Guanacaste dataset			MSCI dataset
Methods	Sey (%)	Spy (%)	TDA (%)	Sey (%)	Spy (%)	TDA (%)
Proposed work (in this article)	98.3	98.3	99.1	99.06	99.34	99.32
Venkatesan Chandran et al. (2021)	92.4	96.2	–	93.20	93.98	94.10
Tripathi et al. (2021)	95.3	96.9	96.8	94.29	94.75	94.28
Dongyao Jia et al. (2020)	93.9	97.1	96.9	93.29	93.67	93.96

Venkatesan Chandran et al. (2021) obtained 93.20% of Sey, 93.98% of Spy and 94.10% of TDA. Tripathi et al. (2021) obtained 94.29% of Sey, 94.75% of Spy and 94.28% of TDA. Dongyao Jia et al. (2020) obtained 93.29% of Sey, 93.67% of Spy and 93.96% of TDA on MSCI dataset. From Table 7, the proposed CEENET methodology obtains high performance than other existing approaches on both Guanacaste and MSCI datasets.

5 Conclusions

An efficient CEENET classification deep learning model is developed in this article to detect and classify the cervical images. Kirsch’s edge detector detects the edge pixels and CWT is used to decompose the edge detected image. The external feature map is constructed from each decomposed sub band and they are trained by developed CEENET model along with the source cervical image. The effectiveness of the CEENET based cervical cancer detection model is tested on Guanacaste and MSCI datasets. In this article, 237 normal cervical images are correctly detected by the proposed system and hence the proposed system achieves 98.7% of NIR for Guanacaste dataset images. Also, 207 cancer cervical images are correctly detected by the proposed system and hence the proposed system achieves 98.5% of NIR for Guanacaste dataset images. Therefore, 98.6% of Mean Index Rate is obtained by the proposed method for Guanacaste dataset images. In this article, 167 normal cervical images are correctly detected by the proposed system and hence the proposed system achieves 98.8% of NIR for MSCI dataset images. Also, 42 cancer cervical images are correctly detected by the proposed system and hence the proposed system achieves 97.6% of NIR for MSCI dataset images. Therefore, 98.2% of Mean Index Rate is obtained by the proposed method on MSCI dataset.

This proposed method is tested on the cancer cervical images and obtained 98.3% of sensitivity, 98.3% of specificity, 99.1% of TDA, 98.1% of PPR and 98.2% of NPR on Guanacaste dataset. This proposed method is tested on the cancer cervical images and obtained 99.06% of sensitivity, 99.34% of specificity, 99.32% of TDA, 98.98% of PPR and 98.96% of NPR on MSCI dataset. The extensive experimental results of the CEENET based cervical cancer detection system provides best cancer region segmentation results in this article. The segmented cancer pixels can be evaluated by different grading levels for clinical evaluation purpose as the future scope of this article.

References

Yusufaly

T.I.

, Kallis

, Simon

, et al. A knowledge-based organ dose prediction tool for brachytherapy treatment planning of patients with cervical cancer, Brachytherapy 19(5) (2020), 624–634.

Zhang

, Luo

Y.M.

, Li

, et al. Cervical precancerous lesions classification using pre-trained densely connected convolutional networks with colposcopy images, Biomedical Signal Processing and Control 55 (2020), 101566.

Stelzle

, Tanaka

L.F.

, Lee

K.K.

, et al. Estimates of the global burden of cervical cancer associated with HIV, The Lancet Global Health 9(2) (2021), e161–e169.

Adem

, Kiliç carslan

and Cömert

, Classification and diagnosis of cervical cancer with stacked autoencoder and softmax classification, Expert Systems with Applications 115 (2019), 557–564.

Matsuo

, Purushotham

, Moeini

, et al. A pilot study in using deep learning to predict limited life expectancy in women with recurrent cervical cancer, American Journal of Obstetrics and Gynecology 217(6) (2017), 703–705.

Shi

, Wang

, Zheng

, Jiang

, Zhang

and Yu

, Cervical cell classification with graph convolutional network, Computer Methods and Programs in Biomedicine 198 (2021), 105807.

Zhang

, Chen

, Gao

, et al. Rapid identification of cervical adenocarcinoma and cervical squamous cell carcinoma tissue based on Raman spectroscopy combined with multiple machine learning algorithms, Photodiagnosis and Photodynamic Therapy 33 (2021), 102104.

, Zhang

, Xin

, et al. Multi-feature based benchmark for cervical dysplasia classification evaluation, Analytical Chemistry Research 63 (2017), 468–475.

Shekar

B.H.

and Uma

K.P.

, Kirsch Directional Derivatives Based Shot Boundary Detection: An Efficient and Accurate Method, Procedia Computer Science 58 (2015), 565–571.

10.

Venmathi

A.R.

, Ganesh

E.N.

and Kumaratharan

, Kirsch Compass Kernel Edge Detection Algorithm for Micro Calcification Clusters in Mammograms, Middle-East Journal of Scientific Research 24(4) (2016), 1530–1535.

11.

Abuobayd, ShabatJules

and Raymond Tapamo , A comparative study of the use of local directional pattern for texture-based informal settlement classification, Journal of Applied Research and Technology 15(3) (2017), 250–258.

12.

Venkatesan Chandran , Sumithra

M.G.

, Alagar Karthick , Tony George , Deivakani

, Balan Elakkiya , Umashankar Subramaniam and Manoharan

, Diagnosis of Cervical Cancer based on Ensemble Deep Learning Network using Colposcopy Images, BioMed Research International 2021 (2021), 1–15.

13.

Ghoneim

, Muhammad

and Hossain

M.S.

, Cervical cancer classification using convolutional neural networks and extreme learning machines, Future Generation Computer Systems 102 (2020), 643–649.

14.

Dongyao Jia

, Zhengyi Li

and Chuanwang Zhang

, Detection of cervical cancer cells based on strong feature CNN-SVM network, Neurocomputing 411 (2020), 112–127.

15.

Tripathi

, Arora

and Bhan

, Classification of cervical cancer using Deep Learning Algorithm, 5th International Conference on Intelligent Computing and Control Systems (ICICCS) 2021 (2021), 1210–1218.

16.

Arora

, Tripathi

and Bhan

, Classification of Cervical Cancer Detection using Machine Learning Algorithms,, 6th International Conference on Inventive Computation Technologies (ICICT) 2021 (2021), 827–835.

17.

http://www.nccc-online.org/index.php/cervicaltumor

18.

, Song

, Ghoneim

and Alrashoud

, Machine learning for assisting cervical cancer diagnosis: an ensemble approach, Future Generation Computer Systems 106 (2020), 199–205.

19.

Saini

S.K.

, Bansal

, Kaur

and Juneja

, ColpoNet for automated cervical cancer screening using colposcopy images, Machine Vision and Applications 31(3) (2020), 1–15.

20.

Adem

and Kiliç

, carslan and O. Cömert, Classification and diagnosis of cervical cancer with stacked autoencoder and softmax classification, Expert Systems with Applications 115 (2019), 557–564.

21.

Karthiga

and Senthil Kumar

, Image Registration based Cervical Cancer Detection and Segmentation Using ANFIS Classifier, Asian Pacific Journal of Cancer Prevention: APJCP 19(11) (2018), 3203–3209.

22.

Chen

, Liu

and Wen

Q.M.

. CytoBrain: Cervical Cancer Screening System Based on Deep Learning Technology, J Comput Sci Technol 36 (2021), 347–360.

23.

https://github.com/yuyao1314/MSCI-dataset

24.

Yao Yu , Jie Ma , Weidong Zhao , Zhenmin Li and Shuai Ding , MSCI: A multistate dataset for colposcopy image classification of cervical cancer screening, Int J Medical Informatics 146 (2021), 104352–104359.

25.

Soumya

M.K.

, Sneha

and Arunvinodh

, Cervical Cancer Detection and Classification Using Texture Analysis, Biomed Pharmacol J 9(2) (2016), 1–10.

26.

Pushpalatha

, Sathiamoorthy

and Kamarasan

, A Novel Framework for Detection of Cervical Cancer, Asian Journal of Engineering and Applied Technology 7(2) (2018), 26–30.

27.

https://www.msci.com/index-data-modules