• <tr id="yyy80"></tr>
  • <sup id="yyy80"></sup>
  • <tfoot id="yyy80"><noscript id="yyy80"></noscript></tfoot>
  • 99热精品在线国产_美女午夜性视频免费_国产精品国产高清国产av_av欧美777_自拍偷自拍亚洲精品老妇_亚洲熟女精品中文字幕_www日本黄色视频网_国产精品野战在线观看 ?

    Automated Grading of Breast Cancer Histopathology Images Using Multilayered Autoencoder

    2022-08-24 03:29:48ShakraMehakUsmanAshrafRabiaZafarAhmedAlghamdiAhmedAlfakeehFawazAlasseryHabibHamamandMuhammadShafiq
    Computers Materials&Continua 2022年5期

    Shakra Mehak,M.Usman Ashraf,Rabia Zafar,Ahmed M.Alghamdi,Ahmed S.Alfakeeh,Fawaz Alassery,Habib Hamam and Muhammad Shafiq

    1Knowledge Unit of System and Technology,University of Management&Technology,Sialkot Campus,51310,Pakistan

    2Department of Computer Science,University of Management and Technology,Sialkot Campus,51310,Pakistan

    3Faculty of Computing and Technology,University of Engineering&Technology,Narowal,51610,Pakistan

    4College of Computer Science and Engineering,University of Jeddah,21493,Saudi Arabia

    5Faculty of Computing and Information Technology,King Abdulaziz University,Jeddah,21589,Saudi Arabia

    6Department of Computer Engineering,College of Computers and Information Technology,Taif University,Taif,Saudi Arabia

    7Faculty of Engineering,Moncton University,NB,E1A3E9,Canada

    8Department of Information and Communication Engineering,Yeungnam University,Gyeongsan,38541,Korea

    Abstract: Breast cancer(BC)is the most widely recognized cancer in women worldwide.By 2018,627,000 women had died of breast cancer(World Health Organization Report 2018).To diagnose BC, the evaluation of tumours is achieved by analysis of histological specimens.At present, the Nottingham Bloom Richardson framework is the least expensive approach used to grade BC aggressiveness.Pathologists contemplate three elements,1.mitotic count,2.gland formation, and 3.nuclear atypia, which is a laborious process that witness’s variations in expert’s opinions.Recently,some algorithms have been proposed for the detection of mitotic cells,but nuclear atypia in breast cancer histopathology has not received much consideration.Nuclear atypia analysis is performed not only to grade BC but also to provide critical information in the discrimination of normal breast,non-invasive breast(usual ductal hyperplasia,atypical ductal hyperplasia)and pre-invasive breast(ductal carcinoma in situ) and invasive breast lesions.We proposed a deep-stacked multi-layer autoencoder ensemble with a softmax layer for the feature extraction and classification process.The classification results show the value of the multilayer autoencoder model in the evaluation of nuclear polymorphisms.The proposed method has indicated promising results, making them more fit in breast cancer grading.

    Keywords: Breast cancer;nuclear atypia;autoencoder

    1 Introduction

    Breast cancer(BC)alludes to a hostile tumour that has been created from nuclei in the breast.BC can originate in three areas:(1)lobules,which are the milk drain organs;(2)ducts,which are paths that empty the drain out of the lobules to the nipples; or (3) stromal tissues, which incorporate the fatty stringy connective tissues of the breast.Among them,the first two are the most common classes of BC[1].Diagnosis with higher precision is desirable to provide treatment efficiently.Regardless of extensive development in the diagnosis and management of breast cancer,it is still the second leading cause of high mortality[2].However,it is necessary to diagnose cancer in the early phases and exact regions to avoid severity.Pathologists use different medical imaging modalities,such as mammography studies,ultrasound scanning, computerized scanning and magnetic resonance imaging, to obtain images of the breast for later analysis[3].The analysis was performed to obtain the aggressiveness of cancerous cells.The aggressiveness of cells is categorically explained by using a grading system that has impacts on treatment.Cancer grading determines how the tumour nuclei look under microscopic tests[4,5].It is not the same as measuring the tumour size and its aggression and metastasis.

    The Nottingham grading system(NGS)proposed by Elston and Ellis is a universal grading system prescribed by the WHO to evaluate BC aggressiveness by pathologists worldwide [6].It compares the appearance of BC tissues with the appearance of normal breast tissues, as shown in Tab.1.It consists of three grades(I),(II),and(III)achieved from the expansion of these three measures:mitotic count,gland formation and nuclear atypia.These three criteria are valued as 1,2,or 3(Tab.1)[6].The evaluation of the BC grade is semi-quantitative.The computation of the three parameters provides tumour aggressiveness.The minimum conceivable score is 3(1+1+1),and the maximum conceivable score is 9(3+3+3).Patients with a score of 8 or 9 are assigned Grade-III.Grade-II alludes to scores of 6 or 7,while Grade-I alludes to scores of 3,4 or 5.High-grade nuclear atypia shows a high deviation in cell shape.Low-grade nuclear atypia,for the most part,has round nuclei[5].

    Table 1:Nottingham grading system by Elston and Ellis

    The rest of this paper is structured as follows.Section 2 outline the latest research on automated image analysis in histopathology and contribution of the paper.Section 3 describes the proposed methodology.Section 4 presents the results and discussion.In the last section, we have drawn our conclusion.

    2 Automated Image Analysis in Histopathology

    The microscopic study of cell morphology plays a vital role for medical experts in decision making[7].Automated nuclei segmentation and classification is a repetitive activity and challenging for pathological images.Due to rapid development in digital pathology,various cancer diagnosis and grading systems have been proposed,including brain image analysis[8–10],cervix[11],lungs[12],liver[13]and breast[5,14–18].In this regard,we can find more systems in[19–22].

    BC image analysis can be applied to handle numerous pathology jobs,such as mitosis detection[18,23],tissue segmentation[24],histological classification or cancer grading[15].Automatic analysis is commonly performed by using hematoxylin and eosin (H&E)-stained slides and deep learning techniques to enhance model performance[25].Breast cancer diagnosis from histopathological images always remains the benchmark in clinical pathology [20–24].Histopathology is the investigation of microscopic variations or abnormalities in tissues.Distinctive types of nuclei in BC histological images call attention to the difficulties in their recognition [26].The images that are obtained through traditional radiology,such as mammograms,ultrasound and MRI,are greyscale and do not cover interesting regions, whereas histopathology images have better quality in diagnosing various diseases, including an extensive variety of cancers, due to distinctive features achieved through the staining process in biopsy tests[27].Cancer detection and grading are performed on histopathological images that are obtained by different high-resolution scanners,such as Aperio XT and Hamamatsu NanoZoomer.Cancer cell identification and grading are achieved at different magnifications, such as 10×, 20× and 40×, because of the tissue structure morphology [28].The obtained whole slide images are exposed to pathologists,who examined the images manually to segment regions of interest for assessing critical information for classification[29].The proper classification of breast lesions can provide critically important diagnostic information to prevent under-and over-treatment and properly guide patient treatment.

    The major problem with the current manual diagnosis system is inter-and intra-observer variability and reproducibility problems.Computer-aided diagnosis(CAD)is a developing interdisciplinary area that consolidatescomponents of machine learning and digital image processing with medicinal information[21].Novel CAD techniques for analysing breast tissue will enable pathologists to have more accurate and reproducible diagnoses, leading to improvements in the ability of pathologists to diagnose and prognosis breast cancer.Several studies exist about breast cancer detection and classification using deep learning.Sadad et al.[4]proposed the methodology to classify pre-segmented breast lumps as cancerous and non-cancerous in X-ray images.They used a convolutional neural network and pre-trained models, combined data augmentation and transfer learning to address overfitting problems.The authors only used limited pre-trained networks on a digital database for screening mammography (DDSM), which might be inadequate to generalize the results of their study.Huynh et al., [19] presented a computer-aided system that extracted 219 breast lesions from breast mammographic images.They used CNN as a feature extractor and support vector machine for classification purposes.Although classification accuracy improved, due to the small number of training data,it might experience overfitting[16].Khan et al.,[21]proposed a framework intended to score cancerous tumour tissues based on ER and PR hormone receptors by using image processing techniques.This system scores ER/PR pervasiveness more accurately than conventional methods that usually assist pathologists in exploring histopathological images.The results are recorded in a database to facilitate further analysis.Doyle et al., [30] segregated poor quality from high-grade BC tumour slides by applying a mix of surface-based components for recognizing tumours from non-cancer,and afterward, extended design elements to recognize low from high evaluations.However, this method depends on the precision of the position of nuclei.In their dataset, the nuclei were commented on physically.However, Abirami et al., [31] proposed an artificial neural network-based classifier to detect micro-classifications in every area of breast cancer mammograms.They used discrete wavelet transform(DWT)for feature extraction and neural networks such as multi-layer perceptron and radial basis function for the classification of BC mammographic images.

    Evaluation of the histopathological images of nuclear polymorphism BC is a qualitative method of treatment for assessment and prognosis.In 2020, a study is conducted for mitotic cell detection by using deep learning techniques instead of image processing techniques, as those techniques are costly and not up to the mark.For detection of mitotic cell Faster-RCNN has been used, but this generates results that has unacceptable rate of false positives.These false positives are generated because of small size of mitotic cell or it is similar looking to normal cell.To reduce false positives post processing is applied.In post processing two deep learning method RESNET 50 and DENSENET-201 has been applied that results in refined set of mitotic cell candidates, which are further refined by using statistical parameters such as shape,texture etc.This technique is tested on two open source databases ICPR 2012,ICPR 2014 and for generalization testing on TUPAC16.On all these datasets the proposed methodology performs well[32].In another study,authors proposed two deep learning models for semantic segmentation and classification to increase robustness.The findings indicate that by combining traditional image processing methodologies and deep learning approaches with hand-crafted features,the efficiency of traditional image processing methodologies and deep learning approaches can be significantly improved[33].

    Different from the existing techniques,the proposed method in our study selects only critical cell nuclei features and utilizes them to grade NA into I,II and III accordingly.The results comprise the prediction of nuclear atypia,which will not merely help pathologists in diagnosis but also bring robust and productive changes in the field of medical imaging.

    To identify histological grades in breast cancer images, we developed an automated grading of nuclei using unsupervised feature extraction.Our major contributions in this paper are as follows:

    ·The proposed system combines a patch-based automated feature extraction method with controlled decision fusion for feature extraction and classification.

    ·An autoencoder-based model is proposed that automatically isolates discriminative patches from high-resolution images for data training.

    ·Our model attains several state-of-the-art outcomes classifying whole slide images(WSIs)into cancer grades and is closely related to inter-observer agreement among pathologists.

    ·Experimental evidence shows that training deep learning models on high-resolution WSI blocks can produce promising results,making them more suitable for breast cancer grading.

    3 Materials and Methods

    3.1 Dataset Description

    In this research, a publicly available nuclear atypia dataset containing Grade-I, II and III BC anomalies by the ICPR (International Conference on Pattern Recognition), is used to assess the effectiveness of the proposed model.The dataset was generated by using 300 manually tagged whole 50 high power field(HPF)slide images of size 1539×1376.These images were scanned by an Aperio scanner at×20 magnification.A solitary WSI of 20×magnification tissue sample comprises almost 1010 multi-channel, representing a highly complex multi-scale structure [34].An HPF is a territory of 0.262 mm2,which is a surface comparable to that of a magnifying lens field breadth of 0.58 mm.The Aperio scanner has a resolution of 0.2273 μm horizontal and 0.222753 μm vertical per pixel.For each frame,two files are maintained.One contains a list of marked scores by each pathologist,and the other is the final score of nuclear atypia decided on majority opinions.The score for nuclear atypia requires a wide zone to have the ability to assess the shape and size of an expansive populace of nuclei.For this purpose,the pathologists worked at×20 magnification,and the final score was given by two or three pathologists.The dataset has been split into training and validation sets.The training and validation sets contain 70 percent and 30 percent of total dataset respectively.Each set contains WSIs for three grades,19%of training dataset is grade-1%,40%is grade-II and remaining is grade-III.In validation set,grade-I,II and III are 16%,42%and 42%respectively.The data used in our experiments also accessible through[35].

    3.2 Data Augmentation

    For deep learning models,the fundamental object is the dataset.Deep learning algorithms truly perform well when the dataset is large[36].This is mainly accurate in the medical imaging field,where access to data is highly protected due to patient secrecy concerns [37].To overcome this matter, we used data augmentation to increase the dataset.The main purpose of data augmentation is generating genuine-looking training data by applying a transformation to an example without altering its label[38–40].Some non-categorical image transformation techniques are used to produce new examples from the training data for image classification,such as cropping,flipping,shifting,and colour jittering,adding Gaussian noise and rotating images at different angles.To balance our dataset,we applied data augmentation of the WSI at different angles(45°,90°,120°and 270°).Algorithms 1 was used for data augmentation.

    3.3 Patch Extraction for Whole Slide Nuclei Image Classification

    Patch extraction is normally used to represent larger scan.Deep neural networks,though,operate on smaller image dimensions as stated in[41]that input size of image may not be larger than 350 by 350 pixels.To reduce the significant image size,input space,computational time and system specifications,WSIs of size 1539×1376 were divided into 128 by 128 size patches,20 manually annotated patches by expert,were used in our study as patch with higher grade will represent WSI[42].As far as number of patches concerns,it is observed that deep neural networks with larger input size may need large number of training parameters i.e.,neurons.This may cause difficulty for neural network during training[41].Every patch consisted of associated features that were representative of the WSI.All patches were assigned a unique identifier and grade according to the Nottingham grade system by a pathologist.Our patch extraction approach is elaborated in Algorithm-2.The overall architecture of the proposed methodology is illustrated in Fig.1.

    Algorithm 1:Data Augmentation Input–Whole slide image(WSI)Output–Augmented whole slide images Step 1:Load WSI I.Step 2:Create image data augmentation generator A (f (i, j), degrees) by a given number of degrees from 0 to 360.Step 3:Generate augmented images.Step 4:Stop.

    Algorithm 2:Generation of patches Input–Whole slide image(WSI)Output–Patches of size n×n of the input image Step 1:Let I be the epidemic patch count and N be the total number of patches.Step 2:Generate patches Step 2.1:Convolve window on the image.Step 2.2:If patch size decreases from n×n,go to step 4.Step 2.3:Extract patch.Step 3:if I=N go to step 4,otherwise go to step 2.Step 4:Stop.

    Figure 1:Framework for nuclei grading based on the Nottingham grading system

    3.4 Feature Learning with Two-Layered Autoencoders

    The architecture of the autoencoder mainly consists of two blocks of layers[34].The first layer is the input layer,which resides as an encoder that is intended to perform the encoding of input data into a lower-dimensional latent compression;a hidden layer can be seen as a new feature representation of input data.The other block contains an output layer that is trained to reconstruct the input from the hidden representation[43,44].The autoencoder simultaneously reduces the divergence between feed data and its reconstruction.The output feature set,which is obtained after applying the transformation function,of every connected layer is the input of successive layers[45].Conventionally,the autoencoder takes input and finds weights that are evaluated on the basis of a threshold.If that threshold matches,data training is stopped; otherwise, weights are again calculated to meet that required criterion.The working architecture of the autoencoder is shown in Fig.2.The autoencoder finds the optimal parameters P= (W,B,B′)by using a hidden layer that executes a transformation function,Eq.(1).

    X= {x1,x2,x3...xn} patches of dimension 128×128 are mapped to dx vectors.The encoder maps the dx vector to an n-dimensional feature vector, i.e., h = ft(x) ∈Vdh.Each neuron in the encoder has a vector of weights that will react to a particular nuclear structure.Later, the features of image patches are passed to the decoder.The decoder tries to change the map to recreate the original input for validation.The high-level feature vector of each patch x, i.e., h ∈V represent cell nuclei atypia.The n feature vector of n patches (X) along with n target values signified asy (m) ∈ {I, II, III},i.e.,{h (m),are then given as input to the softmax layer,which reduces the cost function and acts as a classifier.The list of notations used in modelling is given in Tab.2.We used an autoencoder combined with a softmax layer to shape a deep network.The softmax function can be traditionally used to characterize a categorical distribution that works on the basis of probability function Eq.(2)over different likely results.

    Figure 2:Feature learning process from BC histopathology

    Table 2:Summary of the notations

    In proposed framework,the autoencoder has three hidden layers,the output of the hidden layer is given as input to the next layer.The first hidden layer generates 3000 features which were given as input to the second layer that filter out 300 features from given feature set.Later on,these 300 features maps into 50 features by third layer.The final 50 features are used in classification as shown in Fig.2.

    3.5 Classification

    3.5.1 Support Vector Machine

    The support vector machine(SVM)is a supervised machine learning classifier that combines the influence of conventional theoretical statistical approaches[46].SVMs are commonly used classifiers in various machine learning-based healthcare areas,such as medical imaging[47]and bioinformatics[48].SVM incorporates the capability to overcome the issue of overfitting by using the idea of error minimization.The SVM classifier looks for the ideal hyperplane,which is in the middle of the classes[49].This ideal unscrambling hyperactive plane has numerous refined measurable qualities.SVM additionally learns a non-linear decision function that captures extreme edge hyperplanes by a kernel in high-dimensional vector space that is mapped from input space [50].Typically, SVM addresses noisy data problems and is usually used for binary classification, but it can be specifically stretched out for multi-class classification.In this study,SVM was used for paired characterization of aggressive tumours of Grades I,II and III,as shown in Fig.3.For classification,we designed three parameters for each grade as:xj=(0 0 1),i.e.,Grade-I,b)yj=(0 1 0),i.e.,Grade-II,c)zj=(1 0 0),i.e.,Grade-III.

    Figure 3:Multi-class support vector machine using three hyperplanes,classifying Grades-I,II and III

    3.5.2 Decision Tree

    Numerous machine learning(ML)algorithms are available,among which j48 and random forest are the most frequently used decision trees for classification [51].A decision tree is a method that classifies data as an n-array hierarchy.The initial step,known as tree building,generates a decision tree model through training data and a function, i.e., entropy for obtaining the latest knowledge trends,given in Eq.(3).The other step is tree pruning,in which the insignificant branches do not contribute.To classify,from the root hub,the property estimation of the entered records is tested until a leaf hub is achieved for its respective class.The purpose of tree pruning is to decrease the background distortion or noise in the training data [10,52,53].In this study, random forest (Fig.4) and J48 decision tree(Fig.5)classifiers are used.The random forest(RF)algorithm classifies a given problem by executing the combination of a few decision trees where each tree depends on the value of a distinctly sampled random vector.The distinction of such ensembles of classifiers is that their tree-based components are developed from a specific measure of irregularity,and random forest is characterized as a non-specific standard of the randomized gathering of decision trees.In addition to randomization, achieved by developing trees using bootstrapping, a second layer of randomization is presented at the hub level when growing the tree.RF predicts the output by mapping test features to rules that are generated by using information gain,i.e.,Eq.(4).Finally,it considers the highly voted predicted target as the final prediction from random forest.

    wherepnis the probability that i occurs in the system.A is the target class,and X is the attribute.

    Figure 4:RF architecture that generates N trees to classify nuclei classification

    The J48 decision tree for classification purposes was also used.J48 uses entropy to calculate the homogeneity of the sample data.If the entropy is zero, then the sample data are completely homogeneous, and if the sample data are divided unequally, then it has entropy one.The supplementary features of J48 are detecting missing values,finding persistent characteristic value ranges,determining principles and decision tree pruning.This algorithm delivers the principles for the prediction of the objective variable.In different algorithms, the classification is performed recursively until each sheet is pure; that is, the classification of the data must be as precise as might be expected in the circumstances[54].

    Figure 5:The J48 decision tree shows that the features with the highest information gain classify them into their perspective classes

    3.5.3 Multi-Layer Perceptron Network

    Multi-layer perceptron (MLP) is a supervised neural network consisting of many processing parameters,i.e.,weights that are linked with each other.MLPs are neural systems with an input layer comprised of entities,at least one computation hidden layer that uses the default activation function and an output layer comprising computation nodes.An optimization class is used to minimize the given loss function [55].All attributes of the system are standardized before being given as input.In the input layer,the feature vector is given to the source node that generates the input signal.The input signal moves onward on a path via layers and produces an output signal.As a final point,the computational process terminates on the output layer,which produces results as shown in Fig.6.

    Figure 6:Multi-layer perceptron with N hidden Neurons

    3.5.4 Performance Metrics for Classification

    The quantitative performance of our framework with different models is shown in Tabs.3–7, where we computed our results using metrics in Eqs.(5)–(8).The performance of automatic grading on the basis of nuclei is quantified in terms of the F-measure, precision, and recall(sensitivity)[56,57].

    X = correctly identified instances, Y = incorrectly identified instances, N = number of identified instances that do not align with manual labels.

    Table 3:Results for SVM

    Table 4:Results for RF

    Table 5:Results for J48

    Table 6:Results for MLP

    Table 7:Results for softmax

    4 Results and Discussion

    The proposed framework is evaluated on a publicly available nuclei dataset.This dataset consists of 30 HPF whole slide images at×20 magnification.The study is implemented on a desktop computer(16 GB RAM with Intel Core(TM)3.2 GHz processor)and Intel graphic processing unit.For software implementation,we performed experiments using MATLAB 2016a.

    Deep learning models require a large dataset for effective learning, we augmented the data and extracted them into 128×128 equal size patches to reduce the computational complexity.Our input patch contains significant A101_01= 128×128 =16,384 pixels characterized as learning features.Each patch contains malignant and benign areas of different grades.For feature learning,we used an unsupervised deep learning-based autoencoder model that contains hidden layers along with encoding and decoding layers.The features extracted from a single patch are pixel vectors.This vector is s a portion vector of pixel forces with degree(128×128×3)×1.Thus,there are Sx=S0=128×128×3 input units in the input layer.The first,second and third hidden layers have h(i)=3000,j(i)=300 and K(i)=50,where i=1,2,3,...,n hidden units,respectively.Finally,the softmax layer classifies these 50 neurons into 3 grades accordingly.

    Along with automatic classification with autoencoders,we used various other classifiers because softmax layer classification does not show promising results for this problem.The overall accuracy that was achieved using the softmax function was 63%.The extracted features are given as input to four different machine learning classifiers, i.e.; SVM, RF, J48 and MLP.For SVM, a normalized feature vector is given as input to the RBF kernel with cost 0.1 and precision threshold 0.0005.Second,we use RF for classification with attribute selection and noise removal performed by information gain and pruning,respectively.Along with RF,we also classify grades by considering another decision tree algorithm known as J48.Finally,we use a multi-layer neural network that has 50 hidden neurons and sigmoids as the activation function.The results are compiled at patch-level,however the final label of WSIs can be predicated by considering grade that is assigned to maximum patches of respective WSI.All four classifiers show better results than softmax,as shown in Tabs.3–7.

    Figure 7:Performance of proposed classifies:(a)recall;(b)F-measure,(c)precision;(d)accuracy

    In the classification problem, the sizes of the training and testing data are constrained.We used the N-fold cross-validation technique, where N=10.Various N-fold cross-validations attempt distinctive things with a comparable learning design,and the data index afterwards conveys differing results because of the effect of sporadic assortment in selecting the folds.Stratification reduces the assortment, yet it most likely does not abstain from it totally.We used the N-fold cross-validation method for estimating a tuning parameter t(subset size).This strategy reserves a certain quantity of data for training and testing.N-fold cross-validation is used to predict the error rate of feature learning techniques.We divided our data set into N equal parts,for each N=1,2,3,...,10 fit the model with t parameter to the other N-1 parts,givingα-N(t)and computed its error in predicting the Nth part.Since it was a hectic and tedious job to have a specialist or pathologist that physically recognized every single nucleus in each of these patches(to give a ground truth to quantitative assessment),we compared the results of classifiers with ground truth details provided along with the dataset.The metrics used to evaluate the nuclei scoring of each method include sensitivity,precision,F-measure and accuracy.The performance measures of all classifiers are shown in Fig.7 including(a)recall;(b)F-measure,(c)precision;(d)accuracy.

    5 Conclusions

    The core aim of this study was to explore challenges in the automated nuclei grading framework for H&E images.The qualitative analysis of BC histopathology images was achieved by performing patch extraction,feature learning,and classification of nuclei.The outcome of the proposed methodology was comparable with pathologists’observations given along the dataset.This technique can conceivably help build a computerized grading system for BC.A significant number of studies have been conducted in the field of computer-aided diagnostic tools,emphasizing nuclei segmentation and classification in various image modalities.Nevertheless,there are some open areas to delve into,which may have unique challenges that should be covered in future research.This study was designed to focus specifically on feature engineering using multi-layered autoencoder.Later on in part b of this study,we will test our extracted features using more advance classifiers.Furthermore,we intended to compute the BC grade by combining mitosis and nuclei regions.We also plan to conduct an investigation of the relationship between different grading parameters to evaluate the morphology of the breast in BC.

    Acknowledgement:We deeply acknowledge Taif University for supporting this study through Taif University Researchers Supporting Project Number(TURSP-2020/150),Taif University,Taif,Saudi Arabia.

    Funding Statement:This work was supported by Taif University(in Taif,Saudi Arabia)through the Researchers Supporting Project Number(TURSP-2020/150).

    Conflicts of Interest:The authors declare that they have no conflicts of interest to report regarding the present study.

    十分钟在线观看高清视频www| 亚洲一区中文字幕在线| 一级毛片 在线播放| 欧美xxⅹ黑人| 亚洲第一av免费看| 亚洲综合精品二区| 性少妇av在线| 日韩熟女老妇一区二区性免费视频| 十八禁高潮呻吟视频| 一级毛片我不卡| 777米奇影视久久| 国产女主播在线喷水免费视频网站| 午夜福利视频在线观看免费| 啦啦啦在线免费观看视频4| 国产女主播在线喷水免费视频网站| 久久精品aⅴ一区二区三区四区 | 五月天丁香电影| 亚洲国产av影院在线观看| 黑丝袜美女国产一区| 黑人巨大精品欧美一区二区蜜桃| 黄色怎么调成土黄色| 少妇被粗大猛烈的视频| 亚洲av电影在线进入| 黑人巨大精品欧美一区二区蜜桃| 国产精品久久久久久av不卡| 国产成人91sexporn| 美女脱内裤让男人舔精品视频| 丝袜喷水一区| 国产 精品1| 美女xxoo啪啪120秒动态图| 亚洲一区二区三区欧美精品| 人人妻人人爽人人添夜夜欢视频| 精品国产超薄肉色丝袜足j| 国产成人av激情在线播放| freevideosex欧美| 人妻一区二区av| 麻豆精品久久久久久蜜桃| 在线观看美女被高潮喷水网站| 欧美人与性动交α欧美精品济南到 | av国产精品久久久久影院| 欧美日韩一级在线毛片| 久久久久精品人妻al黑| 亚洲精品国产av蜜桃| 嫩草影院入口| 在线观看三级黄色| 精品福利永久在线观看| 热99久久久久精品小说推荐| 建设人人有责人人尽责人人享有的| 亚洲欧洲精品一区二区精品久久久 | 9色porny在线观看| 成年av动漫网址| a级毛片黄视频| 免费人妻精品一区二区三区视频| 亚洲精品视频女| 久久久久久久亚洲中文字幕| 午夜福利视频精品| 色吧在线观看| 亚洲人成电影观看| 成人免费观看视频高清| 伦理电影大哥的女人| 国产高清不卡午夜福利| 久久久久久久亚洲中文字幕| 五月天丁香电影| 人人澡人人妻人| 菩萨蛮人人尽说江南好唐韦庄| 亚洲欧美一区二区三区黑人 | 国产麻豆69| 男人舔女人的私密视频| 亚洲国产最新在线播放| 日韩中文字幕视频在线看片| 久久99一区二区三区| 久久精品熟女亚洲av麻豆精品| 在线免费观看不下载黄p国产| 看免费av毛片| 亚洲欧洲国产日韩| 午夜日本视频在线| 国产精品av久久久久免费| 亚洲精品美女久久久久99蜜臀 | 爱豆传媒免费全集在线观看| 国产精品无大码| 国产成人精品久久二区二区91 | 一边摸一边做爽爽视频免费| 欧美日韩国产mv在线观看视频| 亚洲情色 制服丝袜| 国产亚洲欧美精品永久| 欧美日韩一级在线毛片| 只有这里有精品99| 久久精品国产亚洲av高清一级| 美女高潮到喷水免费观看| a级片在线免费高清观看视频| 亚洲欧美色中文字幕在线| 午夜精品国产一区二区电影| 久久人人爽人人片av| 99久久精品国产国产毛片| 国产1区2区3区精品| 亚洲情色 制服丝袜| 大码成人一级视频| 亚洲欧美成人综合另类久久久| 久久99蜜桃精品久久| 亚洲精品视频女| 日本免费在线观看一区| 亚洲精品一二三| 少妇精品久久久久久久| 亚洲国产成人一精品久久久| 啦啦啦视频在线资源免费观看| 亚洲欧洲日产国产| 国产成人91sexporn| 少妇被粗大猛烈的视频| 午夜日本视频在线| 欧美日韩一区二区视频在线观看视频在线| 九草在线视频观看| 国产精品国产av在线观看| 男人舔女人的私密视频| 久久久精品区二区三区| 免费在线观看视频国产中文字幕亚洲 | 午夜影院在线不卡| 飞空精品影院首页| 日韩一区二区三区影片| 老司机亚洲免费影院| 亚洲国产最新在线播放| 日本av手机在线免费观看| 夫妻午夜视频| 国产视频首页在线观看| 18禁裸乳无遮挡动漫免费视频| 少妇人妻 视频| 国产精品成人在线| 香蕉丝袜av| 看免费成人av毛片| 宅男免费午夜| 美女中出高潮动态图| 精品人妻一区二区三区麻豆| 另类精品久久| 国产黄色视频一区二区在线观看| 嫩草影院入口| 日韩成人av中文字幕在线观看| 高清av免费在线| 免费不卡的大黄色大毛片视频在线观看| 国产精品偷伦视频观看了| 中文天堂在线官网| 成人免费观看视频高清| 国产精品久久久久久久久免| 一级毛片我不卡| 欧美精品国产亚洲| 亚洲人成电影观看| 欧美黄色片欧美黄色片| 少妇被粗大的猛进出69影院| 精品少妇一区二区三区视频日本电影 | 国产精品人妻久久久影院| 欧美激情极品国产一区二区三区| 久久国产精品男人的天堂亚洲| 黄色 视频免费看| av又黄又爽大尺度在线免费看| 天天躁夜夜躁狠狠久久av| 丝瓜视频免费看黄片| 一本色道久久久久久精品综合| 欧美精品一区二区大全| 波野结衣二区三区在线| 18+在线观看网站| 久久久久久人妻| 在现免费观看毛片| 国产精品 欧美亚洲| 免费播放大片免费观看视频在线观看| 视频在线观看一区二区三区| 久久人妻熟女aⅴ| 国产免费又黄又爽又色| 欧美日韩亚洲高清精品| 午夜91福利影院| 久久久久人妻精品一区果冻| 日韩不卡一区二区三区视频在线| 国产熟女欧美一区二区| 久久这里有精品视频免费| tube8黄色片| 国产无遮挡羞羞视频在线观看| 日韩免费高清中文字幕av| 美女xxoo啪啪120秒动态图| 亚洲av福利一区| 又大又黄又爽视频免费| 国产女主播在线喷水免费视频网站| 性高湖久久久久久久久免费观看| 亚洲精品第二区| 国产一区二区在线观看av| 欧美老熟妇乱子伦牲交| av女优亚洲男人天堂| 亚洲国产av影院在线观看| 一区福利在线观看| 在线 av 中文字幕| 日韩大片免费观看网站| 久久久久国产网址| 在现免费观看毛片| 亚洲天堂av无毛| 精品亚洲成a人片在线观看| 黑人欧美特级aaaaaa片| 永久网站在线| 国产野战对白在线观看| 午夜免费鲁丝| 母亲3免费完整高清在线观看 | 亚洲欧美日韩另类电影网站| 高清欧美精品videossex| 欧美日韩国产mv在线观看视频| 在线观看国产h片| 久久久久国产一级毛片高清牌| 国产麻豆69| 黄片无遮挡物在线观看| 中文天堂在线官网| 国产成人精品久久二区二区91 | 美女主播在线视频| 日韩av在线免费看完整版不卡| 免费观看av网站的网址| 黄频高清免费视频| 夫妻性生交免费视频一级片| 久久久久精品人妻al黑| 久久av网站| 国产av一区二区精品久久| 国产日韩欧美亚洲二区| 国产成人aa在线观看| av一本久久久久| 久久久久精品久久久久真实原创| 狠狠精品人妻久久久久久综合| 高清欧美精品videossex| 精品午夜福利在线看| 精品少妇黑人巨大在线播放| 美女视频免费永久观看网站| 亚洲精品国产av成人精品| 在线观看免费高清a一片| 色婷婷久久久亚洲欧美| 午夜免费鲁丝| 国产视频首页在线观看| 欧美日韩精品成人综合77777| 久久精品久久久久久噜噜老黄| 亚洲精品视频女| 亚洲国产精品一区二区三区在线| 寂寞人妻少妇视频99o| 国产成人午夜福利电影在线观看| 18禁动态无遮挡网站| 少妇熟女欧美另类| 亚洲婷婷狠狠爱综合网| 人妻人人澡人人爽人人| 久久午夜综合久久蜜桃| 日韩精品有码人妻一区| 最新的欧美精品一区二区| 亚洲av国产av综合av卡| 性高湖久久久久久久久免费观看| 亚洲精品国产色婷婷电影| 波多野结衣av一区二区av| av天堂久久9| 国产女主播在线喷水免费视频网站| 最近手机中文字幕大全| 久久久a久久爽久久v久久| 在线精品无人区一区二区三| 亚洲美女搞黄在线观看| 热99久久久久精品小说推荐| 精品一区在线观看国产| 日本午夜av视频| 久久这里有精品视频免费| 乱人伦中国视频| 97精品久久久久久久久久精品| 一二三四中文在线观看免费高清| 国产在视频线精品| 高清欧美精品videossex| 咕卡用的链子| 国产精品久久久久久av不卡| 亚洲综合色网址| 久久精品亚洲av国产电影网| 99香蕉大伊视频| 青草久久国产| 欧美日韩视频精品一区| 中文字幕人妻熟女乱码| 亚洲精品aⅴ在线观看| 91国产中文字幕| 欧美激情极品国产一区二区三区| 国产精品.久久久| 日本黄色日本黄色录像| 亚洲,欧美,日韩| 美女脱内裤让男人舔精品视频| 精品国产乱码久久久久久男人| 日本黄色日本黄色录像| 日韩熟女老妇一区二区性免费视频| 久久久久精品人妻al黑| 免费看不卡的av| 欧美在线黄色| 欧美成人午夜免费资源| 欧美日韩亚洲高清精品| 国产欧美日韩一区二区三区在线| 国产成人aa在线观看| 热99国产精品久久久久久7| 丰满少妇做爰视频| 天美传媒精品一区二区| 我的亚洲天堂| 久久久国产一区二区| 亚洲av.av天堂| 国产激情久久老熟女| av片东京热男人的天堂| 两性夫妻黄色片| 欧美日韩一区二区视频在线观看视频在线| 老汉色av国产亚洲站长工具| 精品人妻在线不人妻| 999久久久国产精品视频| 侵犯人妻中文字幕一二三四区| 亚洲国产成人一精品久久久| 99热全是精品| 日韩一区二区三区影片| 国产日韩欧美亚洲二区| 1024香蕉在线观看| 婷婷色综合大香蕉| 最近最新中文字幕免费大全7| 国产成人a∨麻豆精品| 国产精品 欧美亚洲| www.熟女人妻精品国产| 亚洲精品国产色婷婷电影| 一级毛片我不卡| 日韩欧美精品免费久久| 男人爽女人下面视频在线观看| 狠狠精品人妻久久久久久综合| 国产成人一区二区在线| 国产午夜精品一二区理论片| 色94色欧美一区二区| 国产免费现黄频在线看| 亚洲精品自拍成人| 午夜日本视频在线| 一级毛片 在线播放| 成人影院久久| 免费在线观看视频国产中文字幕亚洲 | 久久国产精品大桥未久av| 精品国产一区二区三区久久久樱花| 五月开心婷婷网| 91国产中文字幕| 制服诱惑二区| 日韩中字成人| 性少妇av在线| 日韩一卡2卡3卡4卡2021年| 久久久国产一区二区| 国产激情久久老熟女| 日韩成人av中文字幕在线观看| 一级毛片电影观看| 最近手机中文字幕大全| 麻豆精品久久久久久蜜桃| 18禁观看日本| 国产乱来视频区| 亚洲婷婷狠狠爱综合网| 亚洲综合色网址| 自拍欧美九色日韩亚洲蝌蚪91| 国产亚洲最大av| 亚洲内射少妇av| 国产精品 欧美亚洲| 亚洲av.av天堂| 久久青草综合色| 欧美日韩国产mv在线观看视频| 国产一级毛片在线| 精品亚洲成a人片在线观看| 国产一区二区三区综合在线观看| 午夜福利在线免费观看网站| 天天操日日干夜夜撸| 国产精品亚洲av一区麻豆 | 免费在线观看完整版高清| 亚洲久久久国产精品| 日韩人妻精品一区2区三区| 国产精品无大码| 日本av免费视频播放| 国产成人免费观看mmmm| 观看美女的网站| 美女国产高潮福利片在线看| 人妻少妇偷人精品九色| 亚洲欧美中文字幕日韩二区| 亚洲国产成人一精品久久久| 亚洲三级黄色毛片| 九草在线视频观看| 91国产中文字幕| 国产极品粉嫩免费观看在线| 精品午夜福利在线看| 一本—道久久a久久精品蜜桃钙片| 免费观看性生交大片5| 丰满少妇做爰视频| 热99国产精品久久久久久7| 欧美成人精品欧美一级黄| 日韩,欧美,国产一区二区三区| 五月开心婷婷网| 成人手机av| 天美传媒精品一区二区| 精品国产国语对白av| 丝袜在线中文字幕| 久久精品国产亚洲av涩爱| 老汉色av国产亚洲站长工具| 久久国产精品大桥未久av| freevideosex欧美| 亚洲成人手机| 国产精品蜜桃在线观看| 午夜福利视频在线观看免费| 精品亚洲成国产av| 成年美女黄网站色视频大全免费| 亚洲精品久久久久久婷婷小说| 各种免费的搞黄视频| 丰满乱子伦码专区| 国产老妇伦熟女老妇高清| 久久精品久久久久久久性| 亚洲国产看品久久| 精品一品国产午夜福利视频| 午夜福利一区二区在线看| 久久久久久久久久久免费av| 欧美精品一区二区大全| 又粗又硬又长又爽又黄的视频| 夫妻午夜视频| 建设人人有责人人尽责人人享有的| 在线观看www视频免费| 久久精品国产a三级三级三级| 天天躁狠狠躁夜夜躁狠狠躁| 久久久欧美国产精品| 久久99一区二区三区| 午夜av观看不卡| 亚洲国产av新网站| 亚洲av福利一区| 一边亲一边摸免费视频| 日本爱情动作片www.在线观看| 下体分泌物呈黄色| 精品国产超薄肉色丝袜足j| 国产精品二区激情视频| 狠狠婷婷综合久久久久久88av| 午夜福利网站1000一区二区三区| 黑人猛操日本美女一级片| 麻豆精品久久久久久蜜桃| 国产精品久久久久久精品电影小说| 国产一区有黄有色的免费视频| 精品亚洲成a人片在线观看| 久久人人97超碰香蕉20202| 精品一区二区三卡| 国产欧美亚洲国产| 日韩伦理黄色片| 欧美精品av麻豆av| 成年人午夜在线观看视频| 久久久久久久久免费视频了| 美女福利国产在线| 在现免费观看毛片| 精品国产国语对白av| 麻豆乱淫一区二区| 国产一级毛片在线| 国产在线一区二区三区精| 久久精品国产a三级三级三级| 精品国产乱码久久久久久男人| 99精国产麻豆久久婷婷| 国产福利在线免费观看视频| 精品午夜福利在线看| 亚洲精品第二区| 国产淫语在线视频| 大陆偷拍与自拍| 成人国产av品久久久| 热re99久久精品国产66热6| 国产精品嫩草影院av在线观看| 少妇 在线观看| 在线观看一区二区三区激情| 性高湖久久久久久久久免费观看| 桃花免费在线播放| 国产免费视频播放在线视频| 天天躁夜夜躁狠狠躁躁| 中文字幕人妻丝袜制服| 五月伊人婷婷丁香| 国产av国产精品国产| videos熟女内射| 老汉色av国产亚洲站长工具| 亚洲精品视频女| 黄片播放在线免费| 久久久久精品性色| 国产精品麻豆人妻色哟哟久久| 日韩一区二区视频免费看| 成人黄色视频免费在线看| 欧美精品国产亚洲| 老司机亚洲免费影院| 考比视频在线观看| 国产免费福利视频在线观看| av免费在线看不卡| 午夜老司机福利剧场| 国产精品香港三级国产av潘金莲 | 99精国产麻豆久久婷婷| 国产精品免费大片| 亚洲四区av| 日韩精品免费视频一区二区三区| 国产极品天堂在线| 啦啦啦在线观看免费高清www| 国产成人免费无遮挡视频| 婷婷成人精品国产| 肉色欧美久久久久久久蜜桃| 亚洲一级一片aⅴ在线观看| 韩国av在线不卡| 成年女人在线观看亚洲视频| 天天躁夜夜躁狠狠久久av| 日日爽夜夜爽网站| 在线亚洲精品国产二区图片欧美| 国产不卡av网站在线观看| 国产成人精品久久久久久| 在线观看美女被高潮喷水网站| 韩国av在线不卡| 国产午夜精品一二区理论片| 欧美成人精品欧美一级黄| av网站在线播放免费| 女人久久www免费人成看片| 大片电影免费在线观看免费| 欧美日韩成人在线一区二区| 欧美精品人与动牲交sv欧美| 成人国产麻豆网| www.精华液| 三上悠亚av全集在线观看| 黄片小视频在线播放| av免费观看日本| 国产精品熟女久久久久浪| 亚洲国产精品国产精品| 美女高潮到喷水免费观看| 欧美亚洲 丝袜 人妻 在线| 在线观看三级黄色| 最近手机中文字幕大全| 纯流量卡能插随身wifi吗| 午夜福利网站1000一区二区三区| √禁漫天堂资源中文www| 97精品久久久久久久久久精品| 叶爱在线成人免费视频播放| 日韩不卡一区二区三区视频在线| 日本欧美国产在线视频| 看免费成人av毛片| 久久久精品国产亚洲av高清涩受| 亚洲 欧美一区二区三区| 成年动漫av网址| 成人毛片a级毛片在线播放| 日韩精品有码人妻一区| 精品第一国产精品| 久久久国产欧美日韩av| 日本-黄色视频高清免费观看| 日韩av免费高清视频| av线在线观看网站| 大话2 男鬼变身卡| 少妇的丰满在线观看| 妹子高潮喷水视频| 日韩熟女老妇一区二区性免费视频| 只有这里有精品99| 女性被躁到高潮视频| 一区福利在线观看| 母亲3免费完整高清在线观看 | 国产精品麻豆人妻色哟哟久久| 亚洲欧美精品自产自拍| 精品国产乱码久久久久久小说| 丝袜脚勾引网站| 午夜福利,免费看| 免费观看a级毛片全部| av女优亚洲男人天堂| 狠狠精品人妻久久久久久综合| 成人亚洲精品一区在线观看| 91久久精品国产一区二区三区| 水蜜桃什么品种好| 亚洲,欧美,日韩| 国产一区亚洲一区在线观看| 精品久久久精品久久久| 黄片大片在线免费观看| 正在播放国产对白刺激| 99久久综合精品五月天人人| 岛国在线观看网站| 男女下面进入的视频免费午夜 | 国产精品国产高清国产av| 高潮久久久久久久久久久不卡| 日韩欧美免费精品| 亚洲狠狠婷婷综合久久图片| 久久精品91无色码中文字幕| 精品无人区乱码1区二区| √禁漫天堂资源中文www| 一区二区三区国产精品乱码| 动漫黄色视频在线观看| 在线视频色国产色| 免费高清视频大片| 国产免费现黄频在线看| 男女做爰动态图高潮gif福利片 | 亚洲精品美女久久av网站| 无遮挡黄片免费观看| 亚洲成人免费电影在线观看| 不卡一级毛片| 色哟哟哟哟哟哟| 国产精品免费视频内射| 国产成人免费无遮挡视频| 精品久久久久久久久久免费视频 | 欧美日韩av久久| 亚洲午夜精品一区,二区,三区| 99国产精品一区二区三区| 91字幕亚洲| 成年人免费黄色播放视频| 麻豆av在线久日| 国产人伦9x9x在线观看| 99国产极品粉嫩在线观看| 国产一区二区三区在线臀色熟女 | 性少妇av在线| 91在线观看av| 亚洲国产看品久久| 美女高潮到喷水免费观看| 很黄的视频免费| 国产激情欧美一区二区| 动漫黄色视频在线观看| 五月开心婷婷网| 一边摸一边抽搐一进一小说| 中文字幕人妻熟女乱码| 啦啦啦在线免费观看视频4| 在线十欧美十亚洲十日本专区| 日本欧美视频一区| 黄色女人牲交| 久久性视频一级片| av国产精品久久久久影院| 精品一区二区三区视频在线观看免费 | 宅男免费午夜| 中国美女看黄片| 亚洲avbb在线观看| 国产精品亚洲一级av第二区| 在线观看免费视频网站a站| 淫秽高清视频在线观看| 色婷婷av一区二区三区视频| 一本大道久久a久久精品| 久久人人精品亚洲av| 中文字幕另类日韩欧美亚洲嫩草| 国产欧美日韩一区二区精品| 成人精品一区二区免费| 男人舔女人的私密视频| 亚洲av日韩精品久久久久久密| 很黄的视频免费| 国产精品永久免费网站|