Research on Application of Classification Model Based on Stack Generalization in Staging of Cervical Tissue Pathological Images

Cervical cancer is a malignant tumor that threatens women’s health and life. Cervical pathology examination, as the gold standard for cervical cancer diagnosis, provides an important basis for the surgical plan and postoperative follow-up strategy for cervical cancer. Cervical biopsy diagnosis includes normal, low-grade squamous intraepithelial lesion (LSIL), high-grade squamous intraepithelial lesion (HSIL) and squamous cell carcinoma (SCC). At present, cervical pathology examination still relies on the doctor’s personal clinical experience and subjective judgment, which is time-consuming and may cause misdiagnosis or missed diagnosis. In addition, the current intelligent classification of cervical pathological images still has disadvantages such as imperfect classification system and low classification accuracy. Therefore, this experiment uses the ResNet50 model of the convolutional neural network as the feature extractor, and selects the K-Nearest Neighbour (KNN), Random Forest (RF), Support Vector Machine (SVM) classifiers in Machine Learning to perform cervical tissue pathological images Discrimination, the accuracy of the classification results were 85.83%, 80.33%, and 86.67%. In order to further improve the accuracy of the model and enhance the applicability and stability of the model, this experiment proposes the Stacked Generalization (SK) classification model. The first-layer base learner of the SK classification model selects CNN-KNN, CNN-RF, CNN-SVM, and the second-layer classifier selects Multilayer Perceptron (MLP). Among them, MLP makes the final result by learning the classification performance of the base learner for label discrimination, the accuracy of the classification model after ensemble learning is 90.00%. In addition, this experiment uses the Synthetic Minority Oversampling Technique (SMOTE) algorithm to amplify the training samples, and the amplified data set has a classification accuracy of 89.17% under the training of the SK classification model. The results show that the SK classification model in this experiment has a high classification ability for cervical histopathological images, and has good generalization ability and robustness.


I. INTRODUCTION
Cervical cancer is one of the fastest growing health problems in the world and the main cause of death among women in developing countries [1]. In developed countries, rapid advances in screening technology and prevention methods have effectively controlled the incidence of cervical cancer The associate editor coordinating the review of this manuscript and approving it for publication was Gustavo Callico . [2].In developing countries, the incidence of cervical cancer is relatively high because of underdeveloped medical services and unbalanced medical resources. For example, the uneven distribution of medical resources in China is particularly obvious in the development of the eastern and western regions and has had a tendency to intensify. More regrettable, Xinjiang has a vast terrain, and differences in the medical environment between the north and the south are even worse [3]. Cervical cancer, thyroid cancer and hydatid disease are the three high-incidence diseases in Xinjiang. Therefore, when cervical lesions can be detected early and treated as an intervention, the economic loss of and health threats to the cervical cancer patient can be reduced.
There are three steps [4] in the diagnosis of cervical cancer. The first step involves TCT liquid-based cytology and HPV testing. When abnormal TCT and HPV results are detected [5], electronic colposcopy plays a key role as a ''bridge''. Colposcopy can be used to identify suspicious lesions and take spot biopsy samples. The obtained pathological tissue is sliced and imaged under the microscope. The pathologist makes a diagnosis based on cervical pathological images. The diagnosis result is the ''gold standard'' for the patient's entire treatment process. In the classification of cervical pathological images, as the severity of the disease worsens, the number of atypical immature cells in the epithelium increases from the bottom to the top of the sample, the nucleoli become larger, and the cytoplasm becomes thick and deepens [6]- [8].
Cervical histopathological images play a decisive role in the diagnosis of cervical cancer. However, the current classification and discrimination of cervical tissue pathological images still rely on the pathologist's personal clinical experience and subjective judgment, and the entire discrimination process has problems such as excessive time consumption and large manual investment. At the same time, fatigue caused by pathologists' work experience and working hours is likely to cause misdiagnosis or missed diagnosis. Therefore, how to effectively use computer-aided diagnosis technology to shorten the pathological diagnosis time and improve the classification accuracy of cervical tissue pathological images has important research value. One study [4] proposed a computer-aided diagnosis with a cervical pathology imaging system, which quantifies and classifies the nuclear structure of ultralarge-scale Cervical Intraepithelial Neoplasia(CIN) images with an accuracy rate of 94.25%. The experiment focused on exploring the pathological images of preSCCous cervical lesions. The degree of discrimination has certain practical value for the training and practice of pathologists, but it lacks a more complete classification system for cervical pathology images. A study [9] proposed extracting the texture feature information of pathological cervix images through the grey level co-occurrence matrix (GLCM). The image is segmented using K-means clustering and a labelcontrolled watershed algorithm, and a support vector machine (SVM) is used for classification. This method has a classification accuracy of 90% for Normal and squamous cell carcinoma (SCC), but the classification accuracy of CIN images is only 70%. Therefore, the classification model of this study is not suitable for sample sets with similar feature vectors, and the classification effect seldom meet the satisfactory clinical requirements. Another study [10] developed a machine vision system using the KS400 macro editing language for 230 images at all levels. The classification scores of the Normal and CIN3 samples in the scoring system reached 98.7%, and the classification scores of the microcytosis and CIN1 samples reached 76.5%. This experimental study lacks a connection between and an evaluation of various Normaland CIN-level images and microcytosis tissue characteristics. The study [11] proposes an automated and localized method based on fusion to evaluate abnormalities in cervical cancer tissues. After using support vector machines and linear discriminant analysis methods to vote on the vertical stage of 61 cases, CIN has the highest classification accuracy rate of 88.5%. The overall classification system of most studies is not perfect, and the discrimination effect for highly similar LSILs and HSILs is poor. Aiming at the imaging characteristics of the abovementioned pathological cervical tissues, a single classification model has low applicability. Therefore, the experiment in this article constructed the Stacked Generalization (SK) classification model, In the experiment, feature extraction is performed through the ResNet50 model of convolutional neural network (CNN), The base learner chooses K-Nearest Neighbour (KNN), Random Forest (RF), Support Vector Machine (SVM), The second layer classifier selects Multilayer Perceptron (MLP).

II. MATERIALS
Cervical Cancer screening for women of childbearing age is mainly divided into three steps. The first step is cervical cytology, called thin-layer liquid-based cytology (TCT), which is currently the main method of screening for early cervical cancer. By observing an exfoliated vaginal cell smear under the microscope, an abnormal cervix is found very early. If the patient's TCT result is diagnosed as suspicious high-grade cancer, the second step of electronic colposcopy is required. Colposcopy involves the use of a microscope to magnify the epidermal tissue of the cervix or genitals with a light source and filter to observe the subtle changes in the cervical epithelium and blood vessels, locate suspicious lesions and determine their severity. If abnormalities are found during colposcopy, then multiple biopsies of suspicious lesions should be performed under the guidance of acetic acid or iodine solution staining and sent for histopathological examination. Histopathological examination as the third step of cervical cancer screening is the ''gold standard'' for the diagnosis of cervical disease in patients.
In this experiment, in the preparation of histopathological cervical slices, the first step is to place the removed tissue block into a pre-prepared 10% formalin solution to denature and coagulate the protein to prevent autolysis or bacterial degradation after cell death and to maintain the original morphological structure of the cell. The second step is to use low-concentration to high-concentration alcohol as a dehydrator to gradually remove the water in the tissue block. The third step is to immerse the transparent tissue block in melted paraffin to cool and solidify into the block. The fourth step is to fix the embedded wax block on a microtome and cut into thin slices. Finally, the tissue slices attached to the glass slide are stained with HE and labeled.
The tissue slice imaging instrument in this experiment was provided by Ningbo Kangfeng Bioinformatics Co., Ltd. (KFBIO), including processing equipment, digital pathology scanning system, and pathology information system. Figure 1 shows four types of cervical pathology images.
In Figure 1, a, b, c, d are Normal, low-grade squamous intraepithelial lesion (LSIL), high-grade squamous intraepithelial lesion (HSIL) and squamous cell carcinoma (SCC). Respectively, from these four cervical pathological images, it can be seen that as the degree of a cervical lesion increases, the epithelial cell arrangement becomes increasingly disordered until the polarity disappears, the cells are significantly heterogeneous, the nucleus is enlarged, the chromatin distribution is uneven, and there is a mitotic phase. When SCC cells penetrate the epithelial basement membrane, dendritic, cordlike, diffuse or mass SCC nests may appear in the interstitium. The magnification of the images collected in this experiment is x200. Due to the presence of cell superposition, abnormal cell nuclei, and various cancer nest shapes in the image, the effect of using conventional cell segmentation techniques is not good. This article will use Resne50 model to extract image features.

A. EXPERIMENT PROCEDURE
The experimental data was provided by the First Affiliated Hospital of Xinjiang Medical University. A total of 480 cervical tissue biopsy imaging data of 84 patients who were in the hospital from September 2018 to October 2019 were collected. Among them, 144 were normal, 108 were low-grade squamous intraepithelial lesions (LSIL), 100 were high-grade squamous intraepithelial lesions (HSIL), and 128 were squamous cell carcinoma (SCC). In order to prevent over-fitting, this experiment divides the data according to all the images of each patient as the smallest set. 75% of the training validation set and 25% of the test set are divided randomly according to each category. Therefore, in normal cervical pathology images, there are 108 training validation sets and 36 testing sets; in LSIL cervical pathology images, there are 81 training validation sets and 27 testing sets; in HSIL cervical pathology images, training There are 75 validation sets and 25 test sets; in the SCC cervical pathology images, there are 96 training validation sets and 32 test sets. Therefore, there are a total of 360 sheets in the training verification set and 120 sheets in the test set. Figure 2 is the general flow chart of this research experiment. The ResNet50 classification model in the convolutional neural network trains and discriminates the input cervical tissue pathological images. In addition, the experiment retained the 2048-dimensional feature vector extracted from the ResNet50 model, on this basis, three classifiers of KNN, RF, and SVM were selected for training and discrimination. In the evaluation of the experimental results of the CNN classification model, CNN-KNN classification model, CNN-RF classification model, and CNN-SVM classification model, this experiment selects CNN-KNN, CNN-RF, CNN-SVM classification model for ensemble learning, the purpose of ensemble learning is to further improve the classification performance and generalization ability of the model.The ensemble learning algorithm selects the stack generalization classification model. Among them, the second-level classifier of the stack generalization classification model makes the final label discrimination by learning the classification performance of the first-level base learner, the second layer classifier of the stack generalization model in this experiment selects MLP. In addition, this experiment did a comparative experiment before and after data enhancement, and the enhancement algorithm selected SMOTE.

B. CNN CLASSIFICATION MODEL
In the digital image classification problem, the complexity of the CNN model directly affects the recognition effect. The more layers that the convolutional neural network has, the richer the features that are extracted from different labels [12]. However, the continuous deepening of the network will cause the problems of gradient explosion and gradient disappearance, and the optimization effect will  [13]. This paper uses the ResNet50 model of convolutional neural network. In order to reduce the time consumption of the model, the high-dimensional pathological tissue slice image needs to be cropped, and the cropped size is (224, 224, 3). In addition, in order to improve the performance of the model, this experiment enhances the data by rotating, horizontal and vertical flipping, brightness and contrast. Figure 3 is a diagram of the ResNet50 network structure. In order to prevent the lack of pixel information, the input cervical tissue pathological image (224, 224, 3) needs to be zero-filled (3 * 3), and the input image after zero-filling passes through the convolutional layer (kernel_size=7 * 7, filters = 64, stride=2 * 2, padding=3) extract the feature vector (112, 112, 64), the channel axis is standardized by BatchNorm and Relu nonlinear activation function, and then through a MaxPooling layer (windows=3 * 3, Stride=2 * 2) Perform maximum pooling and obtain the converged feature map. The network layer structure of steps 2 to 4 are four large blocks respectively. Among them, ConvBlock is used to change the dimension of the network, and IDBlock is used to deepen the network. As the number of network layers increases, the local and global feature vectors (7,7,2048) of the cervical pathology image are extracted and sent to the Average Pooling layer (windows=7 * 7), and the 2048-dimensional feature vector is obtained after Flatten flattening. On this basis, the 2048-dimensional feature vector connect the fully connected layer, and finally output the image prediction label. Table 1 is the experimental environment of this Resnet50. Since the number of images is too small, we will use transfer learning for ResNet50 model. We will use frozen the previous layers until flatten layers.

C. CNN-KNN CLASSIFICATION MODEL
It is worth noting that traditional digital image classification algorithms require manual selection of feature extraction algorithms. This step not only requires investigating the color, texture, and shape feature details of the image itself, but also constantly trying various feature extraction algorithms and adjusting parameters. The whole process takes a long time. The neural network algorithm avoids the complexity of feature extraction and has an excellent performance in the field of digital images [14]. Therefore, the feature vectors of KNN, RF, and SVM classifiers in this experiment use the 2048-dimensional feature vectors extracted by CNN in Section 3.1.
The K-nearest neighbour(KNN) algorithm, as a wellknown statistical method for pattern recognition, occupies an important position in machine learning classification algorithms [15], [16]. The basic idea of the KNN algorithm is that if most of the k most similar samples in the feature space of a sample belong to a certain category, the sample also belongs to this category. In the KNN algorithm, the distance between objects is calculated as an index of dissimilarity between objects. This experiment uses Euclidean distance. As shown in formula (1), Euclidean distance calculates the difference between x individual and y individual, and both x individual and y individual contain n-dimensional features. The greater the distance, the greater the difference between individuals.
KNN only determines the category of the sample to be classified based on the nearest one or a few samples in the decision-making process. Therefore, a too large k value may easily cause under-fitting, and too small a value may easily cause over-fitting. In this experiment, a 10-fold crossvalidation was performed on the divided training validation set to determine the best k value. In the 10-fold crossvalidation process, the training validation set (360 cases) needs to be evenly divided into 10 groups, and each subset of data (36 cases) is used as a validation set, and the remaining 9 groups (324 cases) are used as training Set, this will get 10 models, and use the average of the classification accuracy of the final validation set of these 10 models as the performance index of the classifier under this 10-fold cross-validation. The optimization result is shown in Figure 4. In Figure 4, the abscissa represents the value of k, and the ordinate represents the average value of accuracy after 10 cross-validation corresponding to each determination of a k value. The optimal range of k value in this experiment is 1-30. Experimental results show that when k = 10, the average accuracy of 10-fold cross-validation tends to converge, which is 86.11%. Therefore, in this experiment, the k value is set to 10.

D. CNN-RF CLASSIFICATION MODEL
In 2001, Breiman made a modification on the basis of bagging and proposed a classifier that uses multiple trees to train and predict samples -the random forest algorithm. The algorithm is based on a random method to build decision trees, and  then, these decision trees form a forest. There is no correlation between decision trees. When there is a new sample input, each tree processes it independently. A judgement is made and the classification result of the sample is determined according to the principle of majority [17]- [19]. To construct a random forest classification model, this experiment needs to determine the number of trees in the forest n_estimators. Although the larger the n_estimator, the higher the discrimination accuracy of the classifier, but the memory occupied and the training and prediction time will increase accordingly, and the marginal benefit is decreasing. Therefore, this experiment needs to select the largest n_estimator from the acceptable memory/time. Similarly, this experiment optimizes the best n_estimators value through 10-fold cross-validation on the training and validation set. The optimization result is shown in Figure 5. In Figure 5, the abscissa represents the value of n_estimators, and the ordinate represents the average value of the accuracy after 10-fold cross-validation. The optimization range of n_estimators value in this experiment is 50-1000. It can be found that when n_estimators=150, the 10-fold cross-validation average accuracy rate is the highest, which is 84.75%. Therefore, the value of n_estimators in this experiment is set to 150.

E. CNN-SVM CLASSIFICATION MODEL
Support vector machines have the characteristics of a low generalization error rate and fast classification speed, and they are easy to explain and suitable for small -sample classification problems [20], [21]. The main goal of an SVM is to classify samples mapped to the hyperplane, as shown in equation (2).
where ι 0 / 1 is the loss function, which has the characteristics of non-convexity and discontinuity. In future solutions, functions with good mathematical properties are often used to replace it in a process called a substitution loss. In addition, C is the penalty coefficient. Facing the constrained quadratic linear programming problem in equation (2), this paper uses the Lagrangian multiplier method to solve the problem. The equivalent conversion of the dual problem is given in equation (3).
Equation (3) shows that selecting an appropriate kernel function is very important because it not only can reduce the calculation of the inner product of x T i x j but can also improve the accuracy of classification. In addition, the distribution of the penalty coefficient C and the γ value also affects the number of support vectors and the spatial distribution of the sample map. In this experiment, the optimal C and γ are determined through grid optimization and 10-fold crossvalidation. The optimization results are shown in Figure 6. In Figure 6, the x-axis represents the C value, the optimization range is 0.1-1, the y-axis represents the γ value, and the optimization range is 0.001-0.01, and the z-axis represents the average accuracy after 10-fold cross-validation. It can be found that when C = 0.9 and γ = 0.001, the optimization result is the best, and the average accuracy rate is 88.09%.

F. STACKED GENERALIZATION
Ensemble Learning is a type of machine learning framework that completes learning tasks by constructing and combining multiple learners. Its main frameworks include bagging, boosting and stacking [22]- [24].Among them, in order to achieve better performance than a single classifier, the Stacked Generalization algorithm uses the category probability output of several single classifiers as the feature input of the training model of the next layer. The classification model of the next layer makes the final classification judgment by learning new features. It is worth noting 48984 VOLUME 9, 2021 that the choice of the base learner of the SK classification model needs to meet the differences. In addition, classification models with poor classification effects need to be eliminated, because this type of classification model will have a negative effect on improving the classification ability of the fusion model. In this experiment, the SK classification model is selected. The first-layer base learner selects CNN-KNN, CNN-RF, CNN-SVM, and the second-layer learner selects MLP [25], [26]. As shown in Figure 7, the data labels of this experiment are Normal, LSIL, HSIL, SCC. 360 cases in the training validation set and 120 cases in the test set. In the construction of the stack generalization classification model, the 360 training validation sets are divided into 10 groups in turn. When each subset of data (36 cases) is used as a validation set, the remaining 324 data will be used as the training set, so a total of 10 training verification models were obtained. In the first cross-validation model, 324 training sets are trained, and the four category judgment probabilities of 36 verification sets are output, and the dimensions are (36, 12). Therefore, when these 10 models are trained and verified sequentially, a (360, 4) dimensional vector will be obtained. Different from the 10-fold cross-validation of 3.2, 3.3, and 3.4 to determine the best parameters, the 10-fold cross-validation of the stack generalization classification model aims to preserve the discriminative characteristics of the data by the base classifier. This discriminative feature represents the probability of the base classifier discriminating the data set as a certain category. Since there are 3 base classifiers in this experiment, the feature vector of the new training validation set in the stack generalization model is (360, 12) dimensions in total, which is the feature vector of the new training set. Similarly, in the process of each cross-validation, the test set will be discriminated 10 times, so that 10 sets of (120,4) test set feature vectors will be obtained. By averaging the 10 sets of feature vectors, the new test set feature vector (120,12) will be obtained. In this experiment, the second-level classifier of the stack generalization model selects MLP. The new training verification set feature vector is 360 examples, The layers of MLP are fully connected. The bottom layer is the input layer, which is the new feature vector (360, 12), the middle is the hidden layer, and the final is the output layer. Assuming that the input layer is represented by a vector x, the output of the hidden layer is f (w 1 x+b 1 ),Among them, w 1 is the weight (also called the connection coefficient), b 1 is the bias, the function f in this experiment is the sigmoid function, as shown in formula 3. Its purpose is to directly use numerical optimization methods to learn network parameters. In addition, the hidden layer to the output layer can be regarded as softmax regression, so the output of the output layer is softmax(w 1 x 1 +b 1 ). x 1 represents the output f (w 1 x+b 1 ) of the VOLUME 9, 2021 hidden layer. The determination of the connection weight and bias between the layers of MLP is the core of the optimization problem. This experiment uses gradient descent (SGD) to randomly initialize all parameters and iteratively train.

IV. EXPERIMENTAL RESULTS AND ANALYSIS A. EVALUATION PARAMETERS
Each column of the confusion matrix represents the predicted category, and the total number of each column represents the number of data predicted to be that category; each row represents the true attribution category of the data, and the total number of data in each row represents the number of data instances of that category. Taking the two classification as an example, this experiment evaluates the accuracy, precision, specificity, and sensitivity of the parameters. The F1-score calculation formula is (5)    model not only requires greater differences but also A close and high accuracy rate is required. The 70.83% accuracy rate of the CNN classification model in this experiment will have a negative effect on the final decision-making judgment in the later integrated learning, so it will not be considered. Figure 9 is the confusion matrix of the CNN-KNN classification model for each single classification result of this cervical tissue pathology image ('0' means Normal, '1' means LSIL, '2' means HSIL, and '3' means Cancer), Table 3 is  the calculation result of accuracy, specificity, sensitivity and F1-score. It can be found that the classification effect of the CNN-KNN classification model in HSIL is poor. The reason is that the amount of HSIL and SCC data is small, and the sample size of SCC is greater than that of HSIL. After the value of k is determined to be 10, unbalanced samples can easily lead to misjudgement of HSIL as SCC. Figure 10 is the confusion matrix of the CNN-KNN classification model for each single classification result of this cervical tissue pathology image ('0' means Normal, '1' means LSIL, '2' means HSIL, and '3' means Cancer), Table 4 is the calculation result of accuracy, specificity, sensitivity and F1-score. Similarly, in the case of unbalanced samples, the discriminant results of CNN-RF will tend to have more classification labels in the sample set. Therefore, in this classification experiment, when the CNN-RF classification model is applied to cervical histopathological images with greater feature similarity, the classification results of LSIL and HSIL will be more inclined to the Normal and Cancer categories with more samples in the training validation set. Figure 11 is the confusion matrix of the CNN-KNN classification model for each single classification result of this cervical tissue pathology image ('0' means Normal, '1' means LSIL, '2' means HSIL, and '3' means Cancer), Table 5 is the calculation result of accuracy, specificity, sensitivity   and F1-score. Similar to CNN-KNN and CNN-RF, the classification ability of CNN-SVM makes up for the lack of classification ability of small sample data sets in the deep learning field to a certain extent, but it is difficult to distinguish LSIL and HSIL with small training sample size, its classification and discrimination ability has not achieved satisfactory improvement effects. Figure 12 is the confusion matrix of the CNN-KNN classification model for each single classification result of this cervical tissue pathology image ('0' means Normal, '1' means LSIL, '2' means HSIL, and '3' means Cancer), Table 6 is the calculation result of accuracy, specificity, sensitivity and F1-score. It can be found that the MLP classification model has learned the discriminative ability and the difference of the three base classifiers of CNN-KNN, CNN-RF, and CNN-SVM, and finally has a significant improvement to classify imbalanced data sets.

G. ACCURACY AND TIME CONSUMPTION
The accuracy and time consumption of the base learner and integrated learning built in this experiment are shown in Table 5. And the training time includes the process of optimizing the parameters of each classifier. It can be found that the SK classification n model has a good advantage in the accuracy of the identification of cervical pathological tissues. Compared with the base learner, the time consumption of the model is longer, but considering the importance of clinical diagnosis and treatment plan, the accuracy of computeraided diagnosis is an indicator that we need to consider more.

V. MODEL EVALUATION
It can be seen from Figure 13-17 that the classification effect of the CNN classification model is poor, especially in the classification capabilities of LSIL and HSIL; in addition, compared to the classification of the base learner CNN-KNN, CNN-RF, and CNN-SVM As a result, the stack generalization classification model has further improved the ability to discriminate the same batch of sample instances.

VI. DATA ENHANCEMENT
At the end of the experiment, we used the SMOTE algorithm to enhance the 360 training and verification     Table 8 to Table 11, it can be seen that compared with the three classification models of SMOTE-CNN-KNN, SMOTE-CNN-RF, and SMOTE-CNN-SVM, SMOTE-SK has further improved performance. This is consistent with the experimental conclusions before data enhancement. Through the comparative experiments before and after data enhancement, it can be known that the SK classification model built in this experiment has good robust stability and generalization ability.

VII. CONCLUSION
This experimental research adopts the computer-aided diagnosis mode and establishes the SK classification model based on MLP. Among them, the base learner selects CNN-KNN, CNN-RF, and CNN-SVM classifiers. During the experiment, traditional deep learning methods require high sample data volume, so the training parameters are not very detailed, and the classification effect of new instance samples is poor. Considering the applicability of machine learning to the classification ability of small sample data sets. Retained the 2048-dimensional feature vector extracted by CNN and built KNN, RF, SVM classification models. In order to further improve the classification performance, the SK algorithm in ensemble learning was selected in the experiment. The SK classification algorithm retained the classification of the first-level base learner. The discriminant output probability is imported into the MLP as a new feature, and the MLP classifier makes a final judgment on the new instance sample through the learning ability and difference of the learning base learner. In addition, this experiment uses the SMOTE algorithm to enhance the data set, and the classification results are good when the original classification model parameters are unchanged. The data enhancement comparison experiment shows that the MLP-based SK classification model built in this experiment has good generalization ability and robustness. This study explored the relationship between Normal, LSIL, HSIL, and SCC four categories of cervical pathological tissue images. The proposed SK ensemble classification model algorithm based on MLP has good classification capabilities. This experiment is an applied research, using traditional classification algorithms to apply to the classification of cervical pathological tissue images in this experiment, and achieved good classification and discrimination capabilities, aiming to provide a reference for the intelligentization of future medical diagnosis vision Classification model.  He is currently a Professor with the School of Software, Xinjiang University. His current work is focused upon bioinformatics and artificial intelligence in medical diagnosis. VOLUME 9, 2021