A Critical Analysis of Machine Learning and Deep Learning Methods for Cervical Cancer Screening ()
1. Introduction
The World Health Organization (WHO) says cervical cancer is a major public health concern worldwide, with roughly 570,000 new cases and 311,000 deaths each year [1] . Early cervical cancer identification is crucial for improving patient outcomes because it allows for prompt intervention and treatment. Cervical cancer screening with standard approaches such as Pap smear tests and human papillomavirus (HPV) DNA testing has been beneficial in lowering cervical cancer incidence and death rates [2] . These approaches, however, have limits in terms of accuracy and sensitivity, particularly in low-resource situations [3] . Recent improvements in ml and dl approaches have shown promise in increasing cervical cancer screening accuracy and reliability. Algorithms are used in these approaches to learn patterns and characteristics from vast datasets, which may later be used to name aberrant cells or lesions on the cervix. The application of these methods in cervical cancer screening has the potential to Increase the sensitivity and specificity of screening tests, reduce false negatives, and ultimately improve patient outcomes. In this work, we used ml and dl approaches to evaluate 50 cervical cancer screening trials. We look at the datasets that were used, and the performance indicators that were given, and compare the findings to standard screening approaches. We present a critical assessment of the selected publications’ strengths and weaknesses, as well as the difficulties and potential for future study in this subject. The goal of this study is to present a complete and up-to-date overview and an investigation into the use of ml and dl techniques in the screening of cervical cancer, as well as their potential for improving patient outcomes.
2. Related Works
Yang et al. [1] cover the application of machine learning models, notably multi-layer perceptron and random forests, to the analysis of real-world cervical cancer data. The study tried to simulate cervical cancer detection strategies and assess the accuracy of existing mainstream approaches for diagnosis. Furthermore, the researchers intended to use random forests to assess the significance of numerous risk variables for cervical cancer. The researchers demonstrated a few factors that increase the likelihood of developing cervical cancer, including age, sexual partners, and hormonal contraception, using a training set and a test set with a ratio of 0.75:0.25. The confusion matrix was used by the researchers to assess the efficacy of their algorithms for identifying cervical cancer. The results of the study showed a strong correlation between risk variables and cervical cancer. The study also said that greater data might lead to more targeted research and improved direction for worldwide efforts on cervical cancer prevention and public awareness. Overall, this study proves the utility of machine learning models in finding and predicting cervical cancer risk variables. However, further study with bigger and more diverse datasets is needed to enhance the results’ accuracy and dependability.
On the other hand, the characteristics from the cervical histopathology photos are shown in [2] . After that, a support vector machine (SVM) classifier is trained using the acquired features. Six hundred immunohistochemically stained images of cervical histology are used to assess the proposed architecture. The results of the experiment show that the proposed framework can classify cervical histopathology photos that are well, moderately, and poorly differentiated with good accuracy, sensitivity, and specificity. The proposed system performs better than many cutting-edge methods, such as deep learning and traditional machine learning. Additionally, we conduct extensive experiments to look into how various factors—like the size of the training set, the choice of transfer learning model, and the hyperparameters of the SVM classifier—affect the functionality of the suggested framework. The results of our experiments might be useful for further research in this field. Lastly, the suggested transfer learning framework may effectively categorize photos of cervical histology and be used in hospitals to detect cervical cancer early.
To identify cervical cancer with a high degree of sensitivity and accuracy, Alsmariy et al. address the integration of ml techniques [3] . To reduce predictive performance and encourage incorrect categorization, the research aims to address the issue of an imbalanced dataset. The recommended method uses a voting mechanism that combines three classifiers—Decision Tree, Logistic Regression, and Random Forest—to build a classification model using the UCI cervical cancer risk factors dataset. To solve the imbalance problem and decrease dimensions that do not affect model accuracy, the synthetic minority oversampling method (SMOTE) and principal part analysis (PCA) are applied.
According to Khamparia et al. [4] , this study uses machine learning classifiers in conjunction with transfer learning to propose a deep learning framework for identifying and classifying cervical cancer in Pap smear pictures. Feature extraction from cervical pictures is done in the proposed framework using pre-trained CNN models such as InceptionV3, VGG19, Squeeze Net, and ResNet50. The features are then input into dense and flattened layers for the categorization of healthy and sick cervical cells. For medical diagnosis, the IoHT method is used since it relies less on people and reduces the possibility of human mistake. Based on an evaluation of the proposed framework’s performance using the standard Pap smear Herlev dataset, ResNet50 with a random forest classifier yielded the highest classification rate of 97.89%.
It involves the execution of ml methods in R for analyzing cervical cancer risk conditions, which was covered by B. Nithya et al. [5] . The paper focuses on several feature selection strategies for determining critical traits for cervical cancer prediction and builds classifier models using C5.0, Random Forest, RPART, KNN, and SVM algorithms. Based on the available data, C5.0 and random forest classifiers demonstrated good performance in identifying women who had clinical signs of cervical cancer. According to the findings, using a greatest feature subset and repeated k-fold cross-validation approaches can improve prediction accuracy for cervical cancer detection. The study’s assumptions and limitations are also not properly acknowledged, according to the abstract. Finally, the abstract implies that the findings of the study might be used to predict different forms of gynecological cancer.
Utilizing machine learning models based on behavioral traits and related data, the suggested study covered by Laboni Akter et al. [6] aims to predict cervical cancer. It may be concluded from the results that to reduce the global disparity in cervical cancer incidence, new screening technologies are required. With a 93.33% accuracy rate in cervical cancer prediction, the study includes three ml models: XGBoost, Decision Tree, and Random Forest. Furthermore, emphasized in the paper is the importance of dataset attributes and how they affect the way the classification model is built. According to the study, more research with a larger dataset might aid in enhancing the model’s effectiveness. Overall, the suggested study might help to create efficient cervical cancer screening systems that are proper, inexpensive, and simple to use.
Nicholas Wentzensen et al.’s publication [7] describes CYTOREADER, a cloud-based whole-slide imaging platform that analyzes p16/Ki-67 dual-stained (DS) slides for cervical cancer screening using a deep-learning classifier. Three epidemiological studies of cervical and anal precancers gave platform training on biopsy-based gold standards and comparisons with manual DS and conventional Pap. As compared to manual DS and cytology, AI-based DS had a reduced positive rate, but it retained the same sensitivity and significantly outperformed both in terms of specificity and Pap tests. Comparable results were obtained in two cytology systems and anal cytology, indicating the stability of the platform. The programmed DS assessment eliminates the final subjective component of cervical cancer screening, providing doctors and patients with identical results. Because this method is cloud-based, it may be accessed globally.
An ensemble classification method for accurate cervical cancer detection is provided by the research of Qazi Mudassar Ilyas et al. [8] . This method uses a lengthy voting process to predict the optimal classification results. Many classifiers, such as decision trees, support vector machines, random forests, K-nearest neighbors, naive bayes, multiple perceptions, J48 trees, and logistic regression, were assessed by the researchers. With a prediction accuracy of 94%, the proposed model outperforms single classification methods using the same benchmarked datasets. Health professionals can utilize the study’s findings to refer patients with cervical cancer for more effective therapy by getting a second opinion from a qualified and trustworthy source.
Peng Xue et al.’s paper [9] looks at the worldwide burden of female breast and cervical cancer. It finds that women in low- and middle-income countries (LMICs) are disproportionately affected by these diseases because they have less access to early detection and insufficient treatment. The use of conventional detection methods, such as mammography and ultrasound, is restricted in low- income nations because they need sophisticated infrastructure and training. The autonomous detection of cancer using medical imaging has shown promise for DL, a type of AI. The FDA has given its approval to a small count of DL-based diagnostic clinical diagnostic instruments use, despite the fact few research has demonstrated whether DL models are better or worse. In low- and middle-income countries, the research highlights the need for an accurate and dependable method that requires little training for the early cervix and breast cancer detection.
Using Pap smear histology slides, N. Sompawong et al.’s work [10] aims to screen for cervical cancer using Mask R-CNN. The proposed method produced a 57.8% average precision (map), Accuracy, Sensitivity, and Specificity of 91.7%, 91.7%, and 91.7% for each picture. With Mask R-CNN segmentation, the modified method for categorizing individual cells on the TU dataset test achieved 89.8% accuracy, 72.5% sensitivity, and 94.3% specificity. The study is the first to present an instance segmentation model that uses images from Pap smear slides to automatically diagnose cervical cancer. It may be possible to include the algorithm in medical devices that are used for widespread cervical screening by conducting more studies.
ML for supporting cervical cancer identification: An integrated approach is covered by KMubarak Alrashoud et al. [11] . The cervical pictures undergo preprocessing to eliminate noise, adjust their size, and ensure normalcy. Methods like GLCM and LBP are used to extract features including color, texture, and form. Using the chosen features, four classifiers—SVM, RF, KNN, and NB—are trained and assessed. MV and WV are used in conjunction with the ensemble technique to increase the classification’s overall accuracy.
Adami et al. [12] cover international incidence rates of invasive cervical cancer before cytological screening. The authors used ASRs to estimate the international incidence rates of invasive cervical cancer, comparing ASRs between different regions and countries, and performing a sensitivity analysis to assess the impact of data selection criteria.
F Bray et al. [13] cover 50 years of screening in the Nordic countries: quantifying the effects on cervical cancer incidence. The authors used ASRs to estimate changes in cervical cancer incidence and mortality over time, using statistical models to figure out the relative contribution of screening, changes in risk variables, and treatment improvements.
Cervical Cancer Classification Using Image Processing Approach is covered by Chen et al. [14] . The authors examined image processing methods for cervical cancer classification based on performance measures such as accuracy, sensitivity, and specificity after conducting a thorough assessment of the literature.
[15] have information about CytoBrain method for screening for cervical cancer using deep learning technology. CNN in dl models were employed by the researchers to classify images of cervical cells as normal or abnormal. They employed many metrics, including a ROC curve, precision, sensitivity, and specificity, to assess the performance of the model.
U.K.Lilhore et al. [16] , “Hybrid model for detection of cervical cancer using causal analysis and machine learning techniques”, study develops a machine learning algorithm to analyze cervical cancer risk factors using eight physiological parameters. It analyzes SVM, random forest, decision tree, and Boruta, and Boruta analysis shows that it performs well. The study emphasizes the importance of extracting dysfunctional endogenous components for diagnosis and suggests examining sociodemographic factors. In addition, it recommends the involvement of educational institutions to raise awareness about better healthcare. The findings contribute to better detection and prevention of cervical cancer.
Yi Yin et al. [17] cover Automatic quantification and classification of cervical cancer via adaptive nucleus shape modeling. The authors developed an adaptive nucleus shape model to capture variations in nucleus shape across various stages of cervical cancer and used a SVM classifier to classify the images into dissimilar stages of cervix cancer.
The automated use of convolutional neural networks for the classification of cervical infection from cytological images is covered by Punitha et al. [18] . The authors employed a CNN model to differentiate between normal and abnormal cervical cytological images. Following its training by stochastic gradient descent back propagation, the model was evaluated using metrics such as accuracy, sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve. They compared the proposed classification scheme for cervical cancer with existing state-of-the-art methods.
Mariarputham et al.’s [19] cover Classification of Cervical Cancer using Artificial Neural Networks. The authors collected a dataset of cervical cytological images from the Herlev dataset and preprocessed them by segmenting the nuclei and extracting features such as shape, texture, and color. They developed an ANN model with multiple layers to classify the images into normal or abnormal based on the selected features. The area under the receiver operating characteristic (ROC) curve was used to assess several metrics, such as performance of the model, sensitivity, specificity, and accuracy, after training with stochastic gradient descent in a backpropagation approach.
V. Chandran et al.’s [20] Nominated Texture Based Cervical Cancer Classification is discussed. The Herlev collection provided cervical cytological images, which the authors preprocessed by segmenting the nuclei and extracting textural features. Subsequently built a classification model using a KNN classifier; the success of the model was then assessed using measures that include the area under the receiver operating characteristic (ROC) curve, sensitivity, specificity, and accuracy. This study’s methods for identifying cervical cancer in colposcopy images combine a suggested CYENET model with several convolutional filters with an optimized VGG 19 (TL) model. The CNN model training procedure consists of three steps: data preparation, results of CNN model training and classification. According to [21] , five max-pooling layers, four cross-stream normalization levels, twelve activation layers, and fifteen convolutional layers form the CYENET model.
The technique of the proposed system comprises data preparation and collection, exploratory data analysis (EDA), and machine learning for categorization. The dataset is called “Cervical Cancer Behavior Risk Data Set” and was taken from the University of California, Irvine Machine Learning Repository. Min-max feature scaling is utilized during the preprocessing stage of data. EDA aids in the detection of outliers, data distribution, and feature correlation. Following dataset splitting using the percentage split technique, classification is performed using machine learning classifiers such decision trees, random forests, and XG-Boost in [22] .
The research’s technique included gathering cervical cell imaging data, preprocessing the data, extracting features, constructing a hierarchical modular neural network architecture (HMNNA), training the model, and evaluating it using a variety of metrics. Based on the Bethesda approach, cervical cells were categorized using the suggested HMNNA in [23] .
The method used in the research involved the use of a stacked ensemble algorithm for cervical cancer prediction, along with the application of SMOTE (Synthetic Minority Over-sampling Technique) and RFERF (Recursive Feature Elimination with Random Forest) techniques. The SMOTE technique was used for addressing the class imbalance issue in the dataset, while the RFERF method was used for feature selection to find the most relevant features for cervical cancer prediction. The stacked ensemble algorithm was used for combining the predictions of multiple base models to improve the overall prediction accuracy in [24] .
Ghoneim et al. [25] use three CNN models in their suggested system, one of which is a shallow CNN model with two sets of convolutional and two layers with maximum pooling. It also investigates two advanced CNN models that have been pretrained with millions of photos, Caffe Net and VGG-16 Net. With the use of a training set from the intended database, these pre-trained models are adjusted. The system uses either an ELM-based classifier or an AE-based classifier for classification after the CNN models. Fast learning, simple convergence, and minimal randomization are benefits of the ELM-based classifier, while noise removal and feature extraction are advantages of the AE-based classifier.
This study’s Mugad et al. [26] technique used data mining technologies, notably tree-based algorithms, to accurately show the patients who will get cervical cancer. The SMOTE method was used to address the problem of unbalanced data sets, where cancer patients were disproportionately underrepresented. According to the AUROC curve value, the decision tree algorithm fared better than the choice forest and decision jungle algorithms in terms of prediction accuracy. The study’s accurate results for algorithms such as a Random Forest, a Decision Tree, and a K-Nearest Neighbor algorithm may serve as a model for developing a future cervical cancer healthcare system.
Kurnianingsih et al. [27] proposed method that aims to segment cervical cells from Pap smear images and classify them as normal or abnormal using Mask R-CNN and a VGG-like network. Data preprocessing includes separating original images and masks, data augmentation is applied to increase generalizability, and segmentation involves obtaining bounding boxes, class labels, and confidence scores for objects in the images.
The investigation analyzed the effectiveness of a profound learning neural network algorithm with the Cox proportional hazards (CPH) model on 768 cervical cancer patients in order to predict patient outcomes. The deep-learning model significantly outperformed the CPH model in terms of predictions, especially for progression-free survival (PFS) in FS3. Its mean absolute error was likewise lower. However, both models yielded consistent results for clinicopathologic features associated with survival, with the deep-learning model in [28] adding a few additional significant factors.
This paper Ding et al. [29] using machine learning methods and the TCGA database, a cervical cancer survival prediction model (CCSPM) was created with greater accuracy than earlier models. Ten miRNAs, including those linked to cancer stem cells (CSCs), were used in the model to help divide cervical cancer patients into three survival groups. The difficulties in biological data, such as normalization, feature scaling, and missing value imputation, were addressed by preprocessing techniques.
In [30] , Pap smear images were preprocessed by resizing and augmenting them to prevent overfitting. In this case, a mask-based R-CNN collection method is used for a segment to detect normal and abnormal cervical cells nucleus. Transfer learning technique was employed using pre-trained CNN weights to initialize the Backbone network for training with the histological Pap smear slides.
Three steps are included in the integrated strategy that Tan et al. [31] describe to identify putative gene markers in the biology of cervical cancer: gene expression analysis, meta-analysis of many datasets, feature selection, and machine learning analysis. Using statistical analysis and machine learning techniques, the strategy seeks to identify the most important gene markers in cervical cancer.
The use of machine learning algorithms to find cervical cancer with excellent accuracy and sensitivity is discussed by Fernandes et al. [32] The authors propose a technique for joint dimensionality reduction and classification, which combines supervised and unsupervised components to improve class separability for high-dimensional data. They show the effectiveness of the method on cervical cancer screening.
[33] covered the article presents Compact, a cervical cell classification model developed for large-scale cervical cancer screening. The model uses a smaller version of the VGG system and online data augmentation to improve robustness.
However, Alsmariy et al. [34] cover the document’s discussion of several ensemble methods and classification algorithms, including ensemble Learning, Random Forest, Decision Tree, and Logistic Regression. It also addresses how to handle imbalances in datasets and appropriate feature selection techniques. Through the careful selection of pertinent features, the mitigation of overfitting, and the enhancement of overall classification accuracy, these strategies seek to construct appropriate and efficient models.
Also, The WHO [35] Regional Office for Europe will supply technical help and ease collaboration to achieve cervical cancer elimination in the region. The Regional Multistakeholder Committee will find regional indicators to check progress towards achieving the 2030 global goals for cervical cancer elimination.
WHO [36] has released a revised guide on cervical cancer prevention and control strategies with recommendations on screening, treatment of pre-cancer lesions, and HPV vaccination. The target audience includes health-care providers, managers, and national-level decision-makers.
As explained in [37] Korn et al., the Namibia Cervical Cancer Detection and Therapy Program offers VIA and cryotherapy treatments to women who are HIV-positive (20 - 50) and HIV-negative (25 - 50). Cryotherapy, LLETZ, and thermocoagulation are available treatment options. Data is collected through paper-based health facility registers and analyzed for descriptive statistics.
A baseline analysis of the WHO Global Cervical Cancer Elimination Initiative [38] estimated cervical cancer incidence and mortality rates for 185 countries using GLOBOCAN 2020 database and analyzed time trends and country categorization. R statistical software was used for analysis, and the study was not influenced by funders.
Fontham et al. [39] cover The ACS volunteer Guideline Development Group (GDG) follows a rigorous protocol to develop cancer screening guidelines, managing conflict of interest through transparent disclosure and management processes. For the update of the cervical cancer screening guideline, the GDG used two reports commissioned by the USPSTF as sources of evidence, but also conducted a supplemental literature review and used a microsimulation model to simulate the natural history of the disease.
Wei et al. [40] cover CervDetect is a hybrid machine learning approach that shows risk factors for cervical cancer with an accuracy of 93.6%. The approach could be extended to predict other gynecological cancers and disease entities.
With the goal of accurately estimating cervical cancer survival and site-specific recurrence, Guo et al. [41] cover the work, which introduces a ground-breaking unique method to new methods to AI and ML. Examined were 5112 women from four postsecondary schools. Throughout the SVM, there were 268 fatalities and 343 recurrences. An innovative method of developing prediction models that reliably estimated patients’ survival and site-specific recurrence with CC was made possible by machine learning technology. Using data from 5112 CC patients from four tertiary care hospitals—the largest multicenter cohort ever assembled—they trained and evaluated the model externally. The models may use a range of machine-learning techniques to predict several results simultaneously. In addition to RFS and OS periods, these models contain individual likelihood of general survival, general recurrence, and site-specific recurrence. When several variables were included, machine learning models outperformed traditional logistic or Cox models, the construction of an innovative, intuitive online calculator.
On the other hand, the findings of this paper have significant implications for classification of cervical cancer detection using machine learning algorithms are proved in [42] . Geometric and texture cues were employed to differentiate between normal and malignant cells in Pap smear pictures. For texture feature computation, GLCM was used, followed by PCA for dimension reduction and classification using three types of SVM. With 95% accuracy, the polynomial SVM was discovered to be the most correct. The research was confined to Pap smear pictures, but deep learning techniques might be employed for segmentation and classification in the future.
The ramifications of this work go beyond cervical screening in Western Kazakhstan: the discussion by S. Balmagambetova et al. [43] contrasts azur-eosin staining with liquid-based cytology, or “Cell Scan.” Liquid-based cytology (LBC) and traditional Pap smear performance were compared in Western Kazakhstan in this study, and the findings revealed that LBC was not any more effective than the conventional approach in detecting cancer. The study suggests evaluating LBC’s effects as part of a routine screening procedure. It is advised that existing screening programs be updated to incorporate HPV testing along with triage cytology for positive patients. A small sample size and limited findings translation are two of the research’s shortcomings.
According to [44] , this study represents a substantial advancement in the use of ensemble DL networks for the identification of cervical cancer using colposcopy pictures. The study investigates the use of deep CNN, particularly the VGG 19 (TL) model and the proposed CYENET model, to detect cervical cancer in colposcopy images. The CYENET model has fifteen convolutional layers, twelve activation layers, five max pooling layers, and four cross-channel normalization layers. The three phases of the suggested model are data preprocessing, CNN model training, and classification results. The network description of the CYENET model with variable filter sizes in the identical neural block is shown in the table.
Alyafeai et al. [45] cover the paper presents a groundbreaking novel approach to fully automated deep learning pipeline for cervical cancer classification. The research describes a completely automated deep learning pipeline for finding and categorizing cervical cancer. It features a detecting module for the cervix area with an accuracy of 0.68 and two CNN-based classifiers with AUC values of 0.82 for tumor classification. The suggested pipeline is trained and assessed using Cervi gram pictures, and it features a lightweight design that is right for deployment as a smart device application. Future work will try to improve the quality of Cervi gram pictures and give more precise manual labeling of the cervical region of interest.
Hinsen et al. [46] document represents significant progress in ML methods for cervical cancer diagnosis, using KNN and ANN. The technique described in this paper segments cells saw in a microscope and decides whether they are cancerous using machine learning. The technique uses fuzzy-based segmentation and achieves an accuracy of 88.04% for k-NN and 54% for ANN. The authors suggest that the system can be improved by incorporating other classifiers and by also classifying the stage of cancer. Detecting cancer at an early stage can reduce the risk to the patient’s life.
The outcomes of this work have important ramifications for predicting survival outcomes in cervical cancer: cox models vs. deep-learning models, according to [47] . After looking at 40 features, the researchers divided the 2000-2014 newly diagnosed cervical cancer patients into three feature groups. The deep learning neural network model was compared to other models for predicting both overall survival and progression-free survival. The median absolute error and the correlation value were used for evaluating the performance of the models.
This paper is a significant contribution to automated cervical cancer detection through RGVF segmentation and SVM classification. The research presents a system for automated segmentation and categorization of single cellular Pap smear slides in [48] . The RGVF snake approach is used in the segmentation phase to divide the cell into three areas, while the classification phase employs an SVM-based model with an accuracy of 93.78%, sensitivity of 98.96%, and specificity of 96.69%. The suggested approach aims to increase the efficacy of Pap smear screening.
Jiang et al. [49] cover the approach proposed in this paper has MRI based radiomics approach with deep learning for prediction of vessel invasion in early-stage cervix cancer. Using multi-parametric MRI data, the researchers developed deep learning-based radiomic algorithms to differentiate between early-stage cervix cancer that has invaded vessels and that which has not. The findings show that deep neural network-based radiomics approaches can be used to predict vascular invasion in early-stage cervical cancer prior to surgery. The work used an attention ensemble learning technique to obtain high prediction performance, which has significant promise for future clinical applications. Manual processes, such as handmade segmentation, feature creation, or selection, are not needed. Lin et al. [50] cover the findings of this paper have significant implications of deep learning for fully automated tumor segmentation and extraction of magnetic resonance radiomics features in cervical cancer. The goal of this research was to create an automated tumor segmentation approach for cervical cancer using diffusion weighted (DW) images. The researchers examined 169 MR images and created a U-Net convolutional network for segmentation. The study investigated segmentation performance for various combinations of training input sources, as well as training repeatability. Pearson correlation was used to analyze ADC radiomics. The study revealed that the established approach can properly segment cervical cancer tumors, and ADC radiomics can help with cervical cancer diagnosis.
3. Models or Techniques
Decision Tree, A decision tree is used to display patients as having normal or aberrant cervical cells based on characteristics gathered from cervical images. (Page 10) Random Forest: A categorization technique that uses features gathered from cervical images to determine whether cervical cells are abnormal or normal. Support Vector Machine (SVM) is a machine learning system that classifies cervical cells as normal or diseased based on features extracted from images of the cervical region. An artificial neural network (ANN) is a kind of neural network that uses data from cervical images to categorize cervical cells as normal or pathological. Convolutional Neural Network (CNN): A neural network that classifies cells as normal or abnormal based on images of the cervical region. Deep Belief Network (DBN): A neural network that classifies cervical cells as normal or diseased based on features gleaned from photographs of the cervical region. AdaBoost: Uses features from cervical images to categorize cervical cells as normal or abnormal. Using characteristics gathered from cervical image analysis, the approach known as logistic regression is utilized to categorize cervical cells as normal or abnormal. K-Nearest Neighbors (KNN): A classification technique that classifies cervical cells as normal or pathological based on features extracted from images of the cervical region. Gaussian Mixture Model (GMM): This model uses features extracted from cervical images to categorize cervical cells as normal or abnormal. Feign C-Means (FCM): A cervical cell segmentation method in images. MLP (Multi-Layer Perceptron): This technology uses features taken from cervical images to categorize cervical cells as normal or pathological. Extreme Gradient Boosting (XGBoost): This method uses features gathered from cervical images to categorize cervical cells as normal or abnormal.
4. Results
Fifty studies that addressed the application of DL, ML, or image processing for cervix screening were identified that satisfied our inclusion criteria. A range of methods and strategies were employed in these publications, including segmentation, classification, feature extraction, and computer-aided detection (CAD).
4.1. Machine Learning
Of the fifty articles we looked at, twenty-five of them used machine learning methods to test for cervical cancer. Artificial neural networks, support vector machines and decision trees were the most often utilized machine learning techniques (Table 1).
All things considered; machine learning methods appeared to hold potential for raising cervical cancer screening accuracy. High sensitivity and specificity rates were attained by the application of ANNs and SVMs. Nevertheless, several studies pointed out that the efficiency of these algorithms relied on the caliber of the input data and the settings applied.
4.2. Deep Learning
Of the publications we looked at, twenty-two were devoted to the application of deep learning methods for cervical cancer screening (Table 2). Convolutional neural networks were the most popular deep learning technique (CNNs).
CNNs have shown potential in increasing classification accuracy and decreasing false positive rates in cervical cancer screening. Nonetheless, several studies pointed out how difficult it is to get hold of sizable datasets of cervical image annotations for deep learning algorithm training.
4.3. Image Processing
The application of image processing methods for cervical cancer screening was the subject of three of the publications we examined. Segmentation was the most used image processing technique (Table 3).
![]()
Table 1. Summary of papers using machine learning for cervical cancer screening.
![]()
Table 2. Summary of papers using deep learning for cervical cancer screening.
![]()
Table 3. Summary of papers using image processing for cervical cancer screening.
4.4. Trends and Patterns
Several themes and patterns emerged from the 50 studies we evaluated. To begin, everyone agreed that using machine learning, deep learning, and image processing techniques may increase the accuracy and accessibility of cervical cancer screening. Second, most articles focused on using these approaches in combination with current screening procedures, such as Pap smear and HPV tests. Finally, bigger datasets of annotated cervical pictures were required for training and testing machine learning and deep learning.
4.5. Findings Results
According to the literature evaluation, there is potential for increasing the precision and effectiveness of cervix screening using ML, DL, and image manipulation techniques. Specifically, the use of these techniques has been shown to reduce the number of false positives, improve sensitivity and specificity rates, and detect early-stage cervical cancer. Autoencoder networks and image segmentation were found to be particularly effective approaches, with sensitivity and specificity rates ranging from 0.81 to 0.94 and 0.87 to 0.93, respectively. These methods were also shown to be successful in decreasing false positives and identifying precancerous lesions. The use of active learning and hybrid approaches also showed promising results, with active learning reducing the number of false positives and improving classification accuracy and the hybrid approach detecting early-stage cervical cancer and reducing false positives. There are several possible advantages to cervical cancer screening using deep learning, machine learning, and image processing. These methods can decrease the need for invasive procedures, increase screening efficiency and accuracy, and yield faster findings. Furthermore, the use of these techniques can make screening more accessible to populations with limited access to healthcare. However, there are also potential challenges to using these techniques in clinical practice. The quality of the input data and the specific machine learning or image processing techniques used can greatly impact the accuracy of the results. Additionally, the use of these techniques may require specialized training and resources, which could limit their accessibility to some healthcare providers. Future research should aim to confirm the results found in this literature review in larger and more diverse populations. Additionally, research should focus on perfecting the use of these techniques in clinical practice and showing ways to overcome potential challenges. Lastly, as this will be a key component in determining their viability for broad adoption in healthcare, research should also look at the possible cost-effectiveness of utilizing machine learning, deep learning, and image processing for cervical cancer screening.
A quick summary of the 50 publications’ published years, ML/DL methodology, patient or picture count, and accuracy can be found in Table 4 and Table 5. The papers are arranged according to the ML/DL/IP approach used, which makes it straightforward to compare the results of related methods (Table 4).
The following findings are output based on the types of datasets and source of datasets (Table 5).
Overall, our findings point to the possibility of increasing the precision and effectiveness of cervical cancer screening using deep learning, machine learning, and image processing methods. However, further research is needed to address the limitations found in the studies and to evaluate the clinical utility and cost-effectiveness of these approaches.
According to the below Table 6, we can get the following chart (Figure 1) for better visualization.
![]()
Table 4. An overview of the research on cervical cancer screening employing deep learning, machine learning, and image processing methods.
![]()
Table 5. The kind of dataset and its source, which includes findings from an investigation of the study on the application of Image Processing, Deep Learning, and Machine Learning methods to cervical cancer screening.
![]()
Table 6. Finding the number of applications for machine learning, deep learning, and image processing methods in articles from the literature study on cervical cancer screening.
![]()
Figure 1. Number of uses algorithm of papers.
5. Conclusion
Finally, we have shown how machine learning, deep learning, and image processing techniques may be used to increase the accuracy and efficiency of cervical cancer screening based on our review of 50 studies on the subject. Promising results regarding high sensitivity, specificity, and accuracy in identifying cervical cancer and its precursors have been reported by the research included in this review. Our analysis also revealed several trends across the studies. Convolutional neural networks (CNNs) were the most used technique, followed by support vector machines (SVMs) and generative adversarial networks (GANs). Image pre-processing and feature extraction strategies were discovered to have a substantial influence on algorithm performance. Despite the potential benefits of these methodologies, the research had certain drawbacks, such as small sample sizes, a lack of variety in patient groups, and inadequate validation on independent datasets. These limitations underscore the need for further research to address these gaps and evaluate the performance of these techniques in real-world settings.
6. Future Works
One limitation of this review is that the quality of the studies included varied. Some studies had small sample sizes, which may limit the generalizability of the results. Additionally, some studies did not report sensitivity and specificity rates or did not supply enough detail about the machine learning or image processing techniques used, making it difficult to compare the results across studies. Another limitation is that the studies were conducted in different populations with varying levels of access to healthcare and cervical cancer screening. As such, the results may not be generalizable to all populations. Despite these limitations, the studies reviewed suggest that machine learning, deep learning, and image processing techniques hold promise for improving the accuracy and efficiency of cervical cancer screening. Further research is needed to find the most effective approaches and to confirm the results in larger and more diverse populations. The use of image processing techniques showed promise in improving the visualization of cervical images and helping in the detection of abnormal cells. However, the limited number of papers in this category highlights the need for further research in this area.