An Improved Boykov’s Graph Cut-Based Segmentation Technique for the Efficient Detection of Cervical Cancer

The accurate and reliable derivation of the pap smear cell, which contains cytoplasm and nucleus regions, depends on the segmentation process employed in the cervical cancer detection mechanism. In this paper, an Improved Boykov’s Graph Cut-based Conditional Random Fields and Superpixel imposed Semantic Segmentation Technique (IBGC-CRF-SPSST) is proposed for efficient cervical cancer detection. This proposed IBGC-CRF-SPSST embeds the complete benefits of constraint association among pixels and superpixel edge data for accurate determination of the nuclei and cytoplasmic boundaries so as to ensure efficient differentiation of the healthy and unhealthy cancer cells. Finally, the pixel-level forecasting potential of Conditional Random Fields is included for enhancing the degree of semantic-based segmentation accuracy to a predominant level. The experimental evaluated results of the proposed IBGC-CRF-SPSST aim to produce an accuracy of 99.78%, a mean processing time of 2.18sec, a precision of 96%, a sensitivity of 98.92%, and a specificity of 99.32% value which is determined to be excellent and on par with the existing detection techniques used for investigating cervical cancer.

SECTION I. Introduction

Cancer that occurs in women classified in groups of all ages is the second most common type of cancer, which is considered to be cervical cancer [1]. The worldwide burden of cervical cancer cases and mortality, India alone contributes to about nearly 25.42% and 26.48% [1], respectively. Traditionally, cervical cancer is determined to be prevented by periodic screening in order to identify and treat the precancerous lesions in contrast to most kinds of cancers. Cervical cancer is preventable when the precursor lesions are identified and treated on par with most other kinds of cancer. The large scale of mortality and morbidity imposed by cervical cancer is completely unwarranted since the definite cause of cervical cancer is known, and the disease takes a long duration to develop after the initial infection that happens with high-risk Human Papillomavirus (HPV).

The selection of appropriate cervical cancer detection techniques depends on the resources available, the essentiality of reliable biopsy samples, morbidity, and the cost of the screening method. The process of screening is considered to minimize the mortality and incidence of cervical cancer. However, the development of the Pap smear [1] test by George Papanicalou in the early 1950s improved the mortality of cervical cancer cases in developed countries. The manual screening test is used so that cervical cancer can be identified from the pap smear test, which is used in all clinical trials because the cytology-oriented screening has better results than any other screening method. However, the integration of automated computer-assisted techniques with Pap smear tests is confirmed to be highly suitable for the effective screening of cervical cancer.

In the last decades, uterus cancer has been considered a bog-standard disease in females that affects the genital tract featured by cervical cancer. Cancer which is declared the world’s second most dangerous type of cancer is cervical cancer as per records of the World Health Organization (WHO) since it has caused 2 78, 608 deaths each year [2] including 4, 93, 246 women diagnosed with it in each and every year. Cervical cancer deaths happen most frequently among the age of women who lies between 15 and 44 years [3]. The cervix, called the uterine cervix, is located in the lowest portion of the uterus in women and connects the uterus body (endo-cervix) with the birth canal through the exo-cervix [4].

This cervical region is covered by squamous and glandular cells. Generally, cervical cancer occurs in the cervix core layer, [5] which is the intersection region between the uterus and the vagina through the infection of the Human Papilloma Virus (HPV). HPV is mostly transmitted during sexual intercourse from the male to the female. The two common categories of cervical cancer are Squamous cell carcinoma and adenocarcinoma. In the recent past, it has been identified that 80-90% [6] of cervical cancers are mainly due to squamous cell carcinoma, which originates at the junction part of the Exo and endo-cervix.

According to a survey by the Indian cancer society in 2017, women [7] who had more than three full-term pregnancies and younger women at their first full-time pregnancies are considered to have a higher degree [8] of risk of having cervical cancers. In this context, passive smoking, which dramatically damages the DNA present around the cervical cells, is also considered the second factor for getting the risk of cervical cancer, predominately for women. The most symptoms of cervical cancer are pelvic pain, post-menopausal bleeding [9], and abnormal vaginal bleeding. Women may also experience vaginal discharge with blood, bone fractures around the pelvic region, and urine leakage [10] in the advanced stages of cancer detection. The automated detection of cervical cancer techniques is considered to be highly suitable for the potential identification of abnormalities that occurs in the cell regions [11]. The graph cut-segmentation [12] schemes are considered to be an option for the accurate determination of abnormalities in the nuclei and cytoplasm boundaries [13]. However, the existing methods for [14] detection of cervical cancer techniques were determined to be unsuccessful in estimating nuclei and cytoplasm boundaries under the issue of irregular staining and contrast that are more common in Pap smear tests.

In this proposed paper, an Improved Boykov Graph Cut-based Conditional Random Fields and Super Pixel imposed Semantic Segmentation Technique (IBGC-CRF-SPSST) is used for efficient cervical cancer detection. This proposed IBGC-CRF-SPSST embeds the complete benefits of constraint association among pixels and superpixel edge data for accurate determination of the nuclei and cytoplasmic boundaries so as to ensure efficient identification of the healthy and unhealthy cancer pap smear cells. The proposed IBGC-CRF-SPSST also possesses the merit of facilitating superior boundary adherence and remarkable feature extraction potential where cervical cancer can be detected efficiently. The experimental results and investigation of the proposed HLC-BC-GCST approach are facilitated based on the classification accuracy, specificity, processing time per image, and sensitivity for quantifying its predominance over the benchmarked schemes of the literature.

Cervical cancer has become the second leading cause of female-based cancer-associated deaths after breast cancer, which accounts for the total cancer deaths of about 8% [15] in women and needs to be prevented by detecting precancerous cells at its early stage through the employment of precise screening tests. The Pap smear test is [16] considered to be the most successful screening method due to its potential to minimize the morbidity and mortality of cervical cancer. Cervical cancer-infected patients need to be treated at an early stage once the abnormal cells in the cervix are detected. Traditionally, a Pap smear test is based on microscopic observations [17] with the objective of identifying abnormalities in the morphological structure [18] of cells extracted from the cervical region. However, the microscopic observations in the Pap smear test may be inconsistent due to the subjective variation exhibited by various observers or clinicians. This subjective variation in [19] Pap smear tests necessitates the development and introduction of commercial devices and advanced technologies for the purpose of minimizing the false negative rate under the process of screening for cervical cancer. Further, a number of cervical cancer systems and approaches were propounded for detecting abnormalities through the enhancement of specimen preparation and screening in the dimension of visibility and machine-based detection. However, cost-effective automated screening methodologies that completely replace the judging process of visual systems are not available till now. In this context, [20] computer-aided image processing techniques are determined to be suitable for assisting in the artificial diagnosis of tumours or cell abnormalities. This computer-aided image processing technique is confirmed to minimize the impact of manual intervention and enhance the accuracy and efficiency of the screening process [21] From the past decades, computerized methods that include cell nuclei and cytoplasm detection, classification, and segmentation have been increasingly developed in the medical image analysis field.

At this juncture, segmentation and classification are the two important activities in the process of cervical cell screening, whether semi-automatic or more automatic screening methods are employed in the diagnosis process. In specific, the cell segmentation methods used in the semi-automatic or more automatic screening process are classified into pure nuclei segmentation and integrated nuclei and cytoplasm segmentation. A diversified number of techniques that include active contour, thresholding, level sets, and graph cuts are widely used for facilitating the process of predominant cell segmentation. In particular, Graph cut methods of cervical cancer detection techniques are identified to be potent in the accurate detection and delineation of cell nuclei, cell cytoplasm, or cells under varying staining conditions. These Graph cut methods are identified to resolve the issues of overlapping the disturbing objects existing in the direct vicinity of the process for the detection of cervical cancer.

However, these segmentation techniques based on Graph cut need to be computationally efficient since the complete screening [22] process needs to be executed in an acceptable time period. In addition, the cell nuclei and cytoplasm boundaries need to be accurately detected through the incorporation of a potentially strengthened graph cut-based segmentation scheme since they are quite significant in handling the influence of staining and contrast in Pap smear tests.

The contributions of this paper are stated below.

• A novel, Improved Boykov Graph Cut method (IBGC)is proposed to detect the boundaries of the nucleus and cytoplasm cells for cervical cancer detection.

• Semantic segmentation is used to identify the regions of the pixel between the nucleus and cytoplasm area, where a superpixel imposition method is introduced to accurately differentiate both areas.

• A VGG-based FCN model is proposed for feature extraction from the cervical cells in order to classify the area of the nucleus and cytoplasm cells.

• A CRF-based accurate boundary recovery is proposed for the classification of the boundary area and to identify whether it is a normal or cancerous cell.

The other sections of this proposed work are organized as follows: Section II presents significant review results on recent segmentation schemes based on Graph–cut propounded in the survey with the prediction of their outcome. Section III describes the series of steps included in the implementation of the proposed IBGC-CRF-SPSST method. Section IV briefs the outcomes of the experimental analysis organized for quantifying the proposed IBGC-CRF-SPSST approach potentiality over the existing cervical cancer schemes used for detection in the investigation. Section V summarizes the conclusion of the proposed IBGC-CRF-SPSST method and also discusses the future direction.

SECTION II. Related Works

This section briefs the works that focus on the detection of cervical cancer through an effective segmentation process and is presented with its merits and limitations in Table 1. Initially, the detection of cervical cancer approach using the Fuzzy C Means Clustering Scheme was contributed for segmenting individual pap smear cells into cytoplasm region and nucleus region [23]. Fuzzy C Means Clustering Scheme-based segmentation segmented four categories of cells that are derived from LCH and ERUDIT that are related to the low-grade squamous, normal squamous, carcinoma squamous cell, and high-grade squamous epithelial cells. The classification accuracy of this Fuzzy C Means segmentation method obtained 97.93% and 96.20% under the investigations done with the LCH and ERUDIT datasets. Then, a local and global graph cut approach was proposed for segmenting cervical cells through the enforcement of automation-supported reading approaches that aid in minimizing errors with increased productivity [24]. This local and global graph cut approach increased the rate of sensitivity during the processing of segmenting abnormal cervical cells through the integration of concave points that touches the touching nuclei. The classification accuracy and F-measure of this global and local cut were determined to be 93.21% and 88.46% under the investigations done with the Herlev dataset.

TABLE 1 Related Woks Pros, Cons, Dataset and Accuracy

A Support Vector Machine (SVM) classifier-based scheme was proposed for effective segmentation derivation of the cell contours by considering morphology and textual features for classification [25].

Further, a dual-level cascaded classifier-based segmentation approach was proposed for classifying abnormal epithelial cells from normal epithelial cells from the pap smear cells derived from the cervix part [26]. This dual-level cascaded classifier-based segmentation approach aided in classifying morphology and texture features that are unique to each cell category. Then, a Machine Learning Classifier-based detection approach using the merits of k-Nearest Neighbor, SVM, and Decision Tree for effective segmentation and classification process was introduced [27].

Furthermore, an integrated Binary graph cut and alpha extensible graph cut were proposed predominant detection of cervical cancer cells of pap smear cells [28]. This integrated Graph cut utilized binary Graph cut for extracting foreground pixels, in which Gaussian and Laplace filter-based multiple scale filtering process is imposed. A Graph Search-Based Improved Segmentation Technique (GSBIST) of abnormal

pap smear nuclei was proposed for enhancing the degree of coarse features for accurate definition of cytoplasm and nuclei boundaries [29]. This construction of Graph cuts aided in determining the optimal path in a global manner through the enforcement of dynamic programming. An Automated Cancer detection scheme for Pap Smear Methodology Improvement (ACPSMI) using Ripple Type I transform proposed using texture, shape, and colour features using two-level of classification in cervical dysplasia [30]. Furthermore, Graph Segmentation for Automated Cervical Cancer detection (GSANCC) using graph cut locally was introduced for effective segmentation of cancer was considered for detection [31]. In addition, the authors of this paper also have contributed a Neutrosophic Graph Cut-based Segmentation Scheme (NGCSS) that aided in the predominant processing of cervical images for precise detection of cytoplasm and nuclei boundaries [32]. In this section, the most recent works that focus on the detection of cervical cancer through an effective segmentation process are presented with their merits and limitations.

SECTION III. Proposed Work

The proposed IBGC-CRF-SPSST facilitates the process of predominant segmentation over the boundaries of the pap smear cells based on three potential steps viz., i) Initial Cropping and Bias Correction-based preprocessing of pap

smear images, ii) VCG-based FCN model for Feature Extraction, iii) Boundary Optimization using Boykov Graph Cuts through superpixel estimation, iv) CRF-based accurate boundary Recovery for effective semantic segmentation as depicted in Figure 1.

FIGURE 1.

Proposed IBGC-CRF-SPSST Schemes implementation.

A. Initial Cropping and Bias Correction-Based Preprocessing of Pap Smear Images

This process of Bias Correction-based preprocessing of cervical images is essential since the boundaries of crop

cervical cells are not well defined due to the noise and intensity inhomogeneity features inherent in them. Hence, the Bias Correction-based preprocessing step is incorporated to prevent the problem of noise and intensity inhomogeneity features for an effective and potential definition of the boundaries of cropped cervical cells. In addition, the false positive rate of the proposed IBGC-CRF-SPSST scheme is highly reduced through this preprocessing approach that focuses on improving the layer boundaries pertaining to the cytoplasm and nucleus boundaries of cervical pap smear cells.

B. VGG-Based FCN Model for Feature Extraction

The Visual Geometry Group (VGG) is used for image recognition of cervical cells, where a superpixel imposition method is introduced for the classification of pixels. The VGG 16 consists of 16 layers along with the FCN architecture. It classifies the images for the nucleus and cytoplasm regions. In this phase of the proposed IBGC-CRF-SPSST approach, the Fully Convolutional Networks (FCN) are utilized for achieving a predominant feature extraction process for achieving effective semantic segmentation. This FCN model of feature extraction, in contrast to the existing classic Convolutional Neural Network (CNN), is capable of considering arbitrary-size input images for generating related-sized output in order to produce related outputs. In specific, a model using VGG-based FCN is considered in this proposed IBGC-CRF-SPSST approach due to its excellent performance on par with the existing FCN structures. This VGG-based FCN is an enhanced version of the existing VGG-16, which is constructed by replacing the fully connected layers into convolution layers with the first five layers unchanged. This incorporated VGG-based FCN model aids in greatly minimizing the resolution of the resultant feature map after multiple rounds of pooling and convolution. The resolution of feature maps is minimized by 2,4,8, 16 and 32 iterations in the implementation procedure of the proposed IBGC-CRF-SPSST technique.

C. Boundary Optimization Using Boykov Graph Cuts Through Superpixel Estimation

This phase of the proposed IBGC-CRF-SPSST approach aids in effective superpixel estimation for boundary optimization through the formulation of the best Boykov Graph cut. This Boykov graph cut is a combinatorial graph cut that incurs cost depending on the number of edges derived from the modelled Graph. This Boykov graph cut consists of n-links and t-links, which need to be identified at the segmentation boundary in order to ensure the effectiveness of cervical Pap smear cells’ cervical cells segmentation. Thus, the cost of n-links is utilized for efficient segmentation, such that segmentation with the desired balance between the regional and boundary characteristics is facilitated. This Boykov graph cut comprises an arbitrary collection of pixels or voxels with certain neighbouring systems represented through the collection of unordered pairs related to each neighbourhood element. Further, be the considered binary vector that consists of parameters set to in. Thus, the soft constraints that enforce the regional and boundary characteristics of the Boykov Graph cut-based cost function defined through Equation (1) $$\begin{equation*} E(P_{(i)}) =k\ast R(P_{ (i)}) +B (P_{(i)}) \tag{1}\end{equation*}$$ With the regional and boundary term defined through Equations (2) and (3) $$\begin{align*} _{}R\left ({P_{\left ({i }\right)} }\right)&=\sum \nolimits _{a,b \boldsymbol {\in K}} {R(P_{\left ({i }\right)})} \tag{2}\\ R\left ({P_{\left ({i }\right)} }\right)&= \sum \limits _{a,b\epsilon K} R P_{\left ({i }\right)} \ast \xi _{a\ne b} \tag{3}\end{align*}$$ where $\xi$ indicate the relative significance between the regional and boundary characteristics. Furthermore, the boundary term ${B(P}_{\left ({i }\right)})$ must be represented as the penalty cost based on Equation (4). p(j) is the edge between other vertices and is called edges. $$\begin{equation*} B\left ({P_{\left ({i }\right)-\left ({a,b }\right)} }\right)= \frac {1}{dis(a,b)}\ast \mathrm {exp}\left({-\frac {\left |{ \left |{ I_{p\left ({i }\right)}-I_{p\left ({j }\right)} }\right | }\right |^{2}}{2\sigma ^{2}}}\right) \tag{4}\end{equation*}$$

This boundary term ${B(P}_{\left ({i }\right)})$ embeds the penalty cost that is incurred in representing the boundary features of the segmented image ‘I’ with its positive value. This penalty cost depicted in Equation (4) highlights the discontinuity existing between two pixels, a and b investigated for the classification process. If the value of the penalty cost ${B(P}_{\left ({i }\right)})$ seems to diverge away from zero, then the pixels are with similar intensity and highly optimized. In contrast, the intensity of pixels is different and non-optimized when the value of the penalty cost ${B(P}_{\left ({i }\right)})$ converges toward zero. This estimation of penalty cost in the proposed IBGC-CRF-SPSST approach is achieved through the benefits [33] of Laplacian zero crossing. This method of Laplacian zero-crossing aided in estimating the superpixel that guides the boundary optimization process. A superpixel is considered a collection of pixels that are similar in texture, colour, and location. Thus, the superpixel possesses a distinct value of visual importance with the generalized pixels, even though the superpixel does not possess quantifiable semantic information.

D. CRF-Based Effective Semantic Segmentation for Accurate Boundary Recovery

In this phase of the CRF-based accurate Boundary Recovery process, it is essential for enhancing the segmentation accuracy of weak edge, thin structure, and complex superposition of the boundaries of the nucleus and cytoplasm in order to facilitate an effective cervical cancer detection process. Hence, the model of CRF is utilized in this phase of the proposed IBGC-CRF-SPSST approach for effectively recovering the boundaries of the nucleus and cytoplasm for further Optimization in order to ensure maximum accuracy. The Boykov Graph cuts $G=(V, E)$ with pixel-associated labels are modelled after the derivation of global observations that are generally extracted from the input images considered for the detection process. This derivation of global observations is represented as the input images that consist of pixels. Let ‘R’ be the vector modelled through the utilization of random variables with ‘R₁, R₂,…$\text{R}_{\mathrm {N}}$ ’ representing the label set to each pixel ‘i’. In this context, a CRF follows the Gibbs distribution with pair represented through Equation (5) $$\begin{equation*} P(R=r/I)=\frac {1}{G(I)}e^{-E(r\vert I)} \tag{5}\end{equation*}$$ where G(I) is the partition function with the Gibbs energy E(r)of the pixel-associated labels $r \mathrm {\in }L^{N}$ , further, the CRF model that is fully connected is incorporated using Equation (6). $$\begin{equation*} IE\left ({r }\right)= \sum \nolimits _{i} {w_{i}\left ({r_{i} }\right)+ \sum \nolimits _{i,j} {\mathrm {\delta (}r_{i}, r_{j })}} \tag{6}\end{equation*}$$

Which $w_{i}\left ({r_{i} }\right),\delta \left ({r_{i}, r_{j } }\right)$ corresponds to the pairwise and unary potentials for representing the cost and pixel probability associated with each pixel ‘i’ considering the label at a similar point in time. The initial part of the formula with respect to Gibbs energy is considered with the ri because it is the initial start of the vertex and edges from where the consecutive flow of the Graph takes place to the next. The unary potential used in this proposed IBGC-CRF-SPSST approach is considered as the significant boundary optimization feature map in order to enhance the importance of the CRF model. Furthermore, the pairwise potential aid in modelling the association between neighbouring pixels based on colour similarity weighting. This pairwise potential is enforced based on Equation (7) $$\begin{equation*} \omega _{i} \left ({r_{i}, r_{j} }\right)= \alpha \left ({r_{i}, r_{j} }\right)[w_{1}\beta _{2}+ w_{1}\beta _{2}] \tag{7}\end{equation*}$$

With $\beta _{1}$ and $\beta _{2}$ represented through Equations (8) and (9), respectively $$\begin{align*} \beta _{1}&= \mathrm {exp}\left({-\frac {\left |{ \left |{ P_{p\left ({i }\right)}-P_{p\left ({j }\right)} }\right | }\right |^{2}}{2\sigma _{\lambda }^{2}}}\right) - \frac {\left |{ \left |{ I_{p\left ({i }\right)}-I_{p\left ({j }\right)} }\right | }\right |^{2}}{2\sigma _{\mathrm {\partial }}^{2}}\bigg) \tag{8}\\ \beta _{2}&= \mathrm {exp}\left({-\frac {\left |{ \left |{ P_{p\left ({i }\right)}-P_{p\left ({j }\right)} }\right | }\right |^{2}}{2\sigma _{\gamma }^{2}}}\right) \tag{9}\end{align*}$$

At this juncture, $\beta _{1}$ depends on the pixel colour intensities and pixel positions, while $\beta _{2}$ highly dependent only on the pixel position ($P_{\left ({p(i }\right))}, P_{\left ({p(j }\right))})$ highlights pixel positions and ($I_{\left ({p(i }\right))}, I_{\left ({p(j }\right))})$ representing colour vectors). This proposed IBGC-CRF-SPSST $\mathrm {\alpha (}r_{i,}r_{j })$ approach is assigned to 1 under the condition $r_{i \mathrm {\ne }}r_{j }$ or 0 under the condition $r_{i =}r_{j }a\text{s}$ per the Potts model. In addition, the neighbouring pixels set to each individual label of the image pixels need to be penalized, with each similar pixel motivated for assigning a similar label that deviates in terms of pixel distance set to different labels derived from the input pap smear cells used for cervical cancer detection.

SECTION IV. Experimental Analysis and Investigations Reports

The proposed scheme IBGC-CRF-SPSST is implemented using MATLAB 2018 to compare it with the benchmarked NGCSS, GSANCC, ACPSMI, and GSBIST schemes by incorporating the Herlev dataset. This utilized the Herlev dataset comprised of cervical pap smear [34] cells, as shown in Figure 2 of 917 images which were collected and also predominantly stored over a number of years in the University of Herlev Hospital.

FIGURE 2.

Graphical Representation of cell structure.

The utilized dataset of Herlev consists of seven types of infected pap smear cells which relate to the columnar epithelium, superficial epithelium, columnar epithelium, mild dysplasia, intermediate epithelium, severe dysplasia, and moderate epithelium. In particular, the number of severe dysplasia, carcinoma, mild dysplasia, and moderate dysplasia-infected cervical pap smear cells are 192, 146, 183, and 150, respectively, as shown in Figure 3 and Figure 4.

FIGURE 3.

(From left to right represents the Herlev dataset consisting of mild carcinoma cells where the initial preprocessing is applied detection of the nucleus and cytoplasm with semantic segmentation where the first row describes the initial input image, preprocessed image and bias-corrected image. The second row represents the boundary detection of cytoplasm and nucleus region using IBGC_CRF_SPSST.)

FIGURE 4.

(From left to right represents the Herlev dataset consisting of severe carcinoma cells where the initial preprocessing is applied detection of the nucleus and cytoplasm with semantic segmentation where the first row describes the initial input image, preprocessed image and bias-corrected image. The second row represents the boundary detection of cytoplasm and nucleus region using IBGC_CRF_SPSST).

Further, Figure 5 and Figure 6 portray the predominance of the IBGC-CRF-SPSST proposed scheme compared with the benchmarked NGCSS, GSANCC, ACPSMI, and GSBIST approaches using classification accuracy and specificity by increasing the number of rounds under implementation [35]. The accuracy of the classification by the proposed IBGC-CRF-SPSST method is estimated at 99.78%, which is an improvement of 1.73%, 2.11%, 2.85%, and 3.02% compared to the baseline NGCSS, GSANCC, ACPSMI and GSBIST approaches used for comparison. The specificity of the IBGC-CRF-SPSST proposed scheme is estimated at 99.32%, which is an improvement of 2.12%, 2.34%, 2.93%, and 3.52% compared to the baseline NGCSS, GSANCC, ACPSMI and GSBIST approaches used for investigation.

FIGURE 5.

Classification accuracy-proposed IBGC-CRF-SPSST scheme.

FIGURE 6.

Specificity-proposed IB GC-CRF-SPSST scheme-varying.

SECTION Algorithm 1 IBGC CRF SPSST

1:

procedure IBGC CRF SPSST(cervicalSmearImage)

2:

biasCorrectedImage$\leftarrow$ performBiasCorrection(cervicalSmearImage)

3:

croppedImage$\leftarrow$ crop(biasCorrectedImage)

Select the predominant features from VGG16 and FCN 32

4:

features$\leftarrow$ generateFeatureMap (vgg16, imageToVectorcroppedImage, $\prime$ conv5_3’)

5:

minimizedFeatures$\leftarrow$ getMinimizedFeatures(fcn32, features)

Boundary Optimization with Boykov Cut

6:

H, ${W}\leftarrow$ dim(cervicalSmearImage)

7:

nodeNum, edgeNum, edgeStruct$\leftarrow$ createGraph(H, W)

8:

P_k = calculateUnaryPotentials(minimizedFeatures)

9:

${r}\leftarrow$ calculateRegionVector(initialSegmentation)

10:

pairwisePotentials$\leftarrow$ calculatePairwisePotentials(croppedImage)

11:

optimizedSegmentation$\leftarrow$ GraphCut(nodeNum, edgeNum, Pk, r, pairwisePotentials)

CRF based boundary recovery

12:

while true do

13:

crfLabels$\leftarrow$ associatePixelsWithCRF(croppedImage, optimizedSegmentation)

14:

${r}\leftarrow$ calculateRegionV ectorCRF(crfLabels)

15:

$r_{1}$ , $r_{2}\leftarrow \,\,\sum \nolimits _{\mathbf {i-1}}^{\mathbf {len(r)}} {r\left [{ {i} }\right]\mathbf {==0,}\sum \nolimits _{\mathbf {i-1}}^{\mathbf {len(r)}} {r\left [{ {i} }\right]\mathbf {==1}}}$

16:

if $r_{1} ==0$ & $r_{2} ==0$ then continue

17:

else if $r_{1}$ = $r_{2}$ then

18:

boundaryValues$\leftarrow$ extractBoundaryValues(crfLabels)

19:

semanticSegmentation$\leftarrow$ doSemanticSegmentation (inputImage, boundaryValues)

20:

break

21:

end if

22:

end while

23:

end procedure

Furthermore, Figures 7 and 8 exemplars the predominance of the IBGC-CRF-SPSST proposed scheme compared with the benchmarked NGCSS, GSANCC, ACPSMI, and GSBIST approaches using sensitivity and mean processing time by increasing the number of rounds under implementation. The sensitivity of the proposed IBGC-CRF-SPSST scheme is estimated at 99.78%, which is an improvement of 1.92%, 2.23%, 2.93%, and 3.51% compared to the baseline. The time taken to process a single image of the proposed IBGC-CRF-SPSST is estimated to be 2.18 seconds. This can be minimized by 0.14 seconds, 0.43 seconds, 0.60 seconds, and 0.67 seconds excellent to the baseline.

FIGURE 7.

Sensitivity-proposed IBGC-CRF-SPSST scheme-varying rounds.

FIGURE 8.

Mean processing time -proposed IBGC-CRF-SPSST scheme-varying rounds.

In addition, Figures 9 and 10 unveils the potential of the proposed IBGC-CRF-SPSST scheme compared with the benchmarked NGCSS, GSANCC, ACPSMI, and GSBIST approaches. The precision of the proposed IBGC-CRF-SPSST scheme is estimated at 0.963, which is an improvement of 2.13%, 2.83%, 3.12%, and 3.57% compared to the baseline. The recall value of the IBGC-CRF-SPSST proposed scheme is estimated at 0.974, which is an improvement of 2.35%, 2.87%, 3.26%, and 3.97% compared to the baseline techniques.

FIGURE 9.

Precision value-proposed IBGC-CRF-SPSST scheme.

FIGURE 10.

Recall value-proposed IBGC-CRF-SPSST scheme.

‘The forthcoming Tables 2, 3, and 4 portray the predominance of the proposed IBGC-CRF-SPSST over the reviewed detection schemes of cervical cancer presented under related work using the evaluation metrics of classification accuracy, specificity, sensitivity, mean processing time, recall value, and precision. The results of the proposed IBGC-CRF-SPSST quantified in Table 2 is determined to enhance the accuracy and specificity by a significant margin of 3.32% and 4.82%, superior to the reviewed approaches presented in this related work section.

TABLE 2 Accuracy and Specificity of the Proposed IBGC-CRF-SPSST

TABLE 3 Sensitivity and the Proposed IBGC-CRF-SPSSTmean Processing Time

TABLE 4 Precision and Recall Value of the Proposed IBGC-CRF-SPSST

Similarly, the results of the proposed IBGC-CRF-SPSST quantified in Table 3 prove the enhancement rate of sensitivity and mean processing time by 3.12% and 0.14 seconds, remarkable benchmarked detection approaches of cervical cancer considered for analysis. In addition, results of the proposed IBGC-CRF-SPSST quantified in Table 4 inferred an improvement rate of precision and recall by an excellent margin of 4.21% and 3.94% over the benchmarked cervical cancer detection approaches considered for analysis.

SECTION V. Conclusion

In general, Pap smear tests are inevitable for the potential screening of precancerous and advanced cancerous cervical cells to identify their abnormalities. Computer-assisted Pap smear cell segmentation techniques are identified to be suitable for the precise derivation of cytoplasm and nuclei boundaries in order to confirm abnormalities in cervical cells extracted from women’s cervix. However, most of the Pap smear test integrated and computer-automated segmentation schemes are not able to facilitate predominant performance in regular extraction of cytoplasm and nuclei boundaries due to the issue of a hazy and inconsistent staining process. The proposed IBGC-CRF-SPSST was presented for facilitating efficient cervical cancer detection performed by the semantic web where the cervical pap smear cells detections are performed by the predominant definition of cytoplasm and nuclei boundaries. The experimental results of the proposed IBGC-CRF-SPSST are estimated to be improved by 15% in mean accuracy, 18% in mean processing time, 13% in mean precision, 16% in mean sensitivity, and 19% in average specificity value compared to the cervical cancer detection methods for investigation. Future enhancements of the proposed research can be carried out to enhance and address the detection of cervical cancer with sustained classification accuracy, even when the cervical cells contain a double nucleus. The registration process of high and low magnification images can be included in the proposed cervical cancer detection methods for improving the features of the cell that pertain to texture patterns of the nucleus region. The input Pap smear cell images, may be partitioned into blocks, and each individual block may be processed in a parallel way to maintain significant processing time. The Graph cut segmentation process of the proposed research can be enhanced using deep learning-based semantic segmentation techniques for reliable cytoplasm and nuclei boundary estimation in detection.