Cervical cancer is the fourth most common cancer among women worldwide and is mainly caused by human papillomavirus (HPV) infection, with an estimated 604,000 new cases and 342,000 deaths in 2020 [1]. HPV is transmitted through sexual contact and can be prevented by vaccination and screening. Over 90% of new cases and deaths occur in low- and middle-income countries. Women infected with human immunodeficiency virus (HIV) are six times more likely to develop cervical cancer than women without HIV infection [2]. Cervical cancer can be cured if diagnosed early and treated promptly. Therefore, the World Health Organization (WHO) recommends HPV testing as the primary screening method, followed by treatment of pre-cancerous lesions or referral for further evaluation and management of invasive cancer, and WHO has adopted a global strategy to accelerate the elimination of cervical cancer, which involves reaching the 90-70-90 targets by 2030 [3]. These targets are 90% of girls vaccinated against HPV, 70% of women screened with a high-performance test, and 90% of women with cervical disease treated [3].

The Papanicolaou (PAP) smear is the most widely used screening method for cervical cancer. It involves collecting cells from the cervix and examining them under a microscope to identify any abnormalities (Figure 1). The PAP smear test can detect not only cervical cancer but also precancerous lesions that can be treated before they develop into cancer. The PAP smear test has been proven to be effective in reducing the mortality rate of cervical cancer by 70% [4]. However, the PAP smear test also has some limitations, such as:

• It is time-consuming and labor-intensive, as it requires trained cytotechnologists or pathologists to manually review a large number of slides.
• It is subjective and inconsistent, as different experts may have different interpretations and opinions on the same slide.
• It is prone to human errors, such as misclassification, false negatives, false positives, or missed lesions.
• It has low sensitivity and specificity, as it may fail to detect some subtle or rare abnormalities or may confuse some benign conditions with malignant ones.

To overcome these limitations, AI techniques have been applied to PAP smear diagnosis in recent years. AI techniques can automatically analyze PAP smear images and classify them into normal or abnormal categories, as well as detect the severity and type of lesions. AI techniques can also provide quantitative and objective results that are consistent and reproducible. The incorporation of AI in cervical cell screening presents numerous potential advantages, such as heightened diagnostic precision, improved efficiency, and increased patient comfort. AI algorithms can process and analyze large datasets, identifying patterns and inconsistencies that might elude human detection. Furthermore, AI-facilitated screening could lessen the necessity for invasive procedures, reducing patient discomfort and enhancing outcomes.

2. Performance Metrics, Datasets from Image Patches to Whole Slide Images (WSIs), from Machine Learning (ML) to Vision Transformer

Detecting precancerous lesions for cervical cancer in a timely manner is crucial for effective treatment and prevention. Thus, it is imperative to investigate and scrutinize the diverse techniques and methods used for PAP smear diagnostics to enhance the precision and effectiveness of detection. In 1992, the PAPNET, the first commercial automatic screening system, was approved. However, it was only authorized as a method of re-screening for slides that were initially deemed negative by cytologists [5]. The ThinPrep^® imaging system (Version 1.0, Hologic, Marlborough, UK), which was approved as a commercial screening product in 2004, employs a proprietary algorithm to choose the 22 most concerning fields of view (FOV). This has reduced the workload of pathologists while also increasing the accuracy of the process [6]. The ThinPrep^® imaging system uses liquid-based cytology, which has several advantages over traditional methods. For instance, it is viable to obtain a more representative exemplar of the cervix, which curtails the number of mistaken negatives. The system also eliminates the need for manual fixation and staining, which can introduce variability and artifacts. Recent studies have shown that the ThinPrep^® imaging system is more sensitive than traditional cytology screening. Moreover, it is additionally more duplicable, which boosts its correctness. The software is proficient in identifying subtle modifications in cells that could be a sign of cancer, enhancing its effectiveness. Although the ThinPrep^® imaging system has advantages, it also has limitations. Its implementation can be expensive and specialized training is required for operators. Additionally, certain variables, such as the presence of hemoglobin or mucus, can interfere with the investigation. In the year 2008, the emergence of the FocalPoint GS imaging system marked a significant milestone in the field of cervical cytology. The system was designed to identify 10 FOV of cervical cells most likely to be abnormal, which allowed for the stratification of risk and improvement of efficiency [7]. Although there are advantages to this automated system, certain assessments indicate that its cost-effectiveness is confined and may not be suitable for use in developing countries with low to medium development [8]. Additionally, the technology has weaknesses and depends on the final manual screening process [9]. Thus, analysts continue to examine the implementation of AI technology in cervical cytology, with a focus on improving efficiency. The integration of AI technology would enable the automation of the screening process, thereby alleviating the burden on cytopathologists and elevating the accuracy and efficiency of the outcomes. Furthermore, the adoption of AI would bring about heightened cost-effectiveness throughout the entire procedure, which would be particularly beneficial for countries with limited financial resources, where affordability remains a significant concern. By embracing AI technology, the screening process would undergo a remarkable transformation, characterized by swiftness, precision, and cost-efficiency, thereby widening its accessibility to a larger population. Chankong et al. utilized fuzzy c-means clustering technology to segment single-cell images into the nucleus, cytoplasm, and background, thereby achieving whole-cell segmentation [8]. An investigation explored a segmentation model that utilizes images extracted from a PAP smear slide. The model utilized nucleus localization to differentiate normal and abnormal cells, combined with single-cell classification algorithms. The segmentation model achieved a high level of accuracy and sensitivity, respectively, with 91.7% consisting of Mask-RCNN [10]. Recent years have witnessed a transformation in the methods employed for classification, with the majority of approaches no longer depending solely on texture feature extraction or segmentation. One such novel approach involves segmenting cervical single-cell images into the nucleus, cytoplasm, and background, and then extracting morphological features to enable automatic multi-label classification. The consequences of this strategy have been exceedingly promising, with a precision rate above 93%, suggesting the potential efficacy of this methodology in automated sorting [11]. Another innovative approach involves the extraction of seven groups of texture features of cervical cells for classification, with the support vector machine (SVM) classifier demonstrating the highest accuracy and best performance. This strategy is extremely efficient at classifying the pictures with a high level of precision. However, it is worth noting that the precision percentage of the incorporated categorizer was solely 50% at the stain plane and 60% at the unit plane, which indicates the need for further refinement and optimization [12]. Researchers are finding automated categorization methods that do not rely on an accurate segmentation algorithm to be more and more attractive. One of these methods utilizes deep learning (DL) and transfer learning to classify cervical cells. The likelihood of achieving such exceptional performance through manual extraction of deep-level features from cell images for classification, with an accuracy of 98.3%, an AUC of 0.99, and a specificity of 98.3%, is low. This highlights the significant potential of utilizing advanced machine learning techniques, such as deep learning, for improving the accuracy and efficiency of cell image analysis in various medical applications [13]. The complexity and specificity of the task at hand, as described earlier, necessitated the use of a graph convolution network for the precise classification of cervical cells. This advanced machine learning technique achieved impressive results, with precision, sensitivity, specificity, and F-measure rates of 98.37%, 99.80%, 99.60%, and 99.80%, respectively. These findings demonstrate the potential of utilizing graph convolution networks for accurate and efficient analysis of complex medical images [14]. The examination directed by Bao et al. comprises a possible companion examination of a broad populace of females, involving 700,000 people who were going through screening for cervical carcinoma. The AI-assisted cytological diagnostic system employed in the study was validated, resulting in a total coincidence rate of 94.7%. Moreover, this synchronized with a marked upsurge in sensitivity of 5.8% (3.0% to 8.6%) in comparison to manual check. This study demonstrates that the integration of AI-assisted cytological examination can significantly improve the detection and classification of cervical cells and should be considered as a potential tool for guiding triage in cervical cancer screening programs [15]. Zhu and colleagues developed an AIATBS diagnostic system that utilized ThinPrep^® and artificial intelligence and showed higher sensitivity than the diagnosis conducted by experienced cytologists. In fact, the AIATBS system had a remarkable sensitivity of 94.74% when detecting CIN. These discoveries have noteworthy ramifications for the domain of cervical ailment diagnosis, as they propose that AI has the potential to significantly enhance the accuracy and sensitivity of existing diagnostic techniques [16]. Chen and his colleagues carried out research on CytoBrain, a screening system for cervical cancer that employs artificial intelligence. This system utilizes deep learning technology and comprises cervical cell segmentation, classification, and human-aided diagnosis visualization modules. The study mainly focuses on cell segmentation and classification components and proposes a compact VGG network called CompactVGG as the classifier. The researchers introduced a large dataset of 198,952 cervical cell images from 2312 participants, which were categorized into positive (abnormal), negative (normal), and junk categories. The CompactVGG structure features 10 convolutional layers, 4 max pooling layers, and 2 fully connected layers, totaling 1,128,387 parameters. The independent test group found evidence of CompactVGG’s accuracy being 88.30%, sensitivity being 92.83%, specificity being 91.03%, precision being 82.26%, and F1-score being 87.04%. These outcomes surpassed the Inception v3, ResNet50, and DenseNet121 models on all metrics. Furthermore, CompactVGG demonstrated superior time and classification performance compared with existing VGG networks on the Herlev and SIPaKMeD public datasets. In conclusion, the CytoBrain system with its CompactVGG classifier has the potential to improve cervical cancer screening through its accurate and efficient performance [17]. Wei and colleagues introduced an innovative module called InCNet, which enhances the multi-scale connectivity of the network while maintaining efficiency. This module is seamlessly integrated into a lightweight model named YOLCO (You Only Look Cytopathology Once) and is specifically designed to extract features from individual cells or clusters. To evaluate their approach, the authors curated a novel dataset comprising 2019 whole slide images (WSIs) obtained from four different scanners. The dataset includes annotations for both normal and abnormal cells and is publicly accessible for research purposes. In order to assess the performance of their method, the authors compare it with a conventional model that employs a ResNet classifier. The evaluation was conducted on 500 test WSIs. The results demonstrate that the proposed method outperforms the conventional model across most metrics, except for specificity, where the conventional model exhibits a slight advantage. The AUC score achieved by the new method is 0.872, while the conventional model obtains a score of 0.845. Moreover, the accuracy of the proposed method reaches 0.836, surpassing the accuracy of 0.824 achieved by the conventional model. Furthermore, the authors showcase the clinical relevance of their method by illustrating its ability to detect sparse and minute lesion cells in cervical slides. This capability is particularly challenging for human experts and conventional models. The authors assert that their method has the potential to enhance the diagnosis and screening of cervical cancer, thereby contributing to improved healthcare outcomes [18]. Cheng and colleagues developed an innovative system for cervical cancer screening that utilizes deep learning techniques on WSIs. This computer-aided approach has the potential to significantly improve the accuracy and efficiency of cervical cancer screening, offering a promising new avenue for early detection and treatment. The system consists of three models, namely, low-resolution lesion localization, high-resolution cell identification/ranking, and recurrent neural network WSI classification models. To evaluate the efficacy of their system, they used 3545 WSIs from five hospitals and scanners, with 79,911 annotations. In independent testing conducted on 1170 WSIs, the system demonstrated 93.5% specificity and 95.1% sensitivity, similar to three experienced cytopathologists. The system also identified the top 10 lesion cells with a true positive rate of 88.5% on 447 positive slides, surpassing the Hologic ThinPrep^® Imaging System. The computational efficiency of the system is remarkable, given its ability to process giga-pixel WSIs in around 1.5 min per graphic processing unit (GPU), which illustrates its effectiveness in actual screening scenarios [19]. Wang and colleagues proposed an innovative approach for detecting cervical high-grade squamous intraepithelial lesions and squamous cell carcinoma screening in PAP smear WSIs using cascaded fully convolutional networks. This ingenious method involves a step-by-step process of utilizing fully convolutional networks to accurately identify and classify abnormalities in WSIs, which could have significant implications for improving cervical cancer screening and diagnosis. Their investigation, released in a top-tier medical journal, evaluates their proposed deep learning screening system’s efficacy against other state-of-the-art approaches like U-Net, SegNet, and a previous method. The creators shared that their suggested technique obtained a precision of 0.93, recall of 0.90, F-measure of 0.88, and Jaccard index of 0.84 on a 143-WSI dataset, demonstrating remarkable performance over the other methods. Additionally, the proposed method demonstrated a remarkable processing speed of 210 s per WSI, which is 20 times and 19 times faster than U-Net and SegNet, respectively. The potential for an AI system to produce these exceptional results and suggest a method for quickly and accurately identifying severe cervical pathologies in real-world clinical environments could be highly advantageous for the medical community, particularly in areas with limited resources. The results of this analysis demand additional inquiry and authentication in larger and more varied datasets [20]. Kanavati and team designed a model for detecting cervical cancer in liquid cytology WSIs that utilized deep learning, consisting of trained convolutional and recurrent neural networks. The model was tested on 1605 training and 1468 multi-test set WSIs and achieved an ROC AUC range of 0.89–0.96, indicating its potential to assist in screening. Furthermore, the model generates neoplastic cell probability heatmaps that help identify suspicious regions. The model exhibited either comparable or superior accuracy, sensitivity, and specificity when compared with semi-automated devices. Therefore, it has the potential to standardize screening and reduce fatigue. The results of this study suggest that incorporating deep learning into cervical cancer screening could have a substantial impact on the accuracy and efficiency of the screening process. Furthermore, the integration of this model into healthcare may lead to the premature detection of cervical cancer and potentially rescue lives [21]. Hamdi and colleagues developed a novel approach for the analysis of whole slide cervical images and cancer staging using a hybrid deep learning system to generate a combination of models that includes ResNet50, VGG19, GoogLeNet, Random Forest, and support vector machine. The team also utilized the Active Contour Algorithm for segmentation and fused deep model features as another approach. The ResNet50-VGG19-Random Forest model achieved outstanding results on a dataset of 962 cervical squamous cell images. Particularly, the prototype accomplished 97.4% sensitiveness, 99% exactness, 99.6% exactitude, 99.2% selectivity, and 98.75% AUC, which displays noteworthy potential for beforehand detection. Considering the encouraging outcomes of this research, it is probable that upcoming studies may comprise clinical data and more comprehensive, diverse datasets, as well as an investigation of further deep learning models. Overall, this proposed plan of action has the potential to advance the field of cervical cancer diagnosis and improve patient outcomes [22]. Diniz and colleagues conducted a study that involved comparing ten deep convolutional neural networks for classifying cervical cells in PAP smears into two, three, and six categories using conventional cytology images. They proposed an ensemble of the top three models and used extensive data augmentation and balancing. The appraisal of the investigation utilized accuracy, specificity, F1-score, cross-validation, recall, and precision. The ensemble model outperformed individual and prior architectures across all classification tasks. The recall achieved for the two, three, and six classes were 0.96, 0.94, and 0.85. The study highlights the potential of ensemble deep learning in improving the accuracy of cervical cancer screening and, subsequently, patient outcomes. The findings of the investigation suggest that the utilization of an assembly model may result in improved efficacy when compared with individual models. The authors further established the efficacy of enhancing data through augmentation and balancing to enhance the precision of the model. This study provides valuable insights into the use of advanced ensemble deep learning in cervical cancer screening and has the potential to inspire further research in this field [23]. Tripathi and colleagues conducted a study that involved the classification of five cervical cancer cell types in 966 PAP smear images using four pre-trained deep models. These architectures were constructed by a team of highly skilled engineers and were carefully tested to ensure optimal performance. ResNet-152 reached the maximum precision of 94.89%, with VGG-19, ResNet-50, and VGG-16 close on its heels. Furthermore, the study reported class-wise performance, and certain combinations of models achieved 100% recall and precision for specific classes. The investigation accentuates the potential of deep transfer learning for precise classification and implies that further progressions in original models, hyperparameter optimization, and clinical data integration could boost the accuracy even further. In line with the data, this analysis provides valuable insights for future research in the field of cervical cancer grouping. These findings can contribute to improving cervical cancer diagnosis and patient outcomes [24]. Zhou and colleagues presented a comprehensive cervical screening framework that includes three stages: cell detection, image classification, and case classification. The first stage involved detecting cells using the RetinaNet model, which achieved an impressive 0.715 average precision in just 0.128 s per image. The following step integrated an innovative patch encoder-fusion component for image classification, achieving a 0.921 accuracy and 0.903 sensitivity. For the final phase, a support vector machine functioned as a case classifier, providing an accuracy of 0.905 and sensitivity of 0.891, outclassing other models. These numerical results clearly demonstrate the framework’s effectiveness in leveraging cell cues to improve the robustness of case diagnosis. In general, this inventive system shows vast potential for enhancing the detection and diagnosis of cervical cancer, ultimately resulting in improved health outcomes for women. Additional exploration is warranted to affirm the framework’s effectiveness in larger and more diverse patient populations [25]. The CervixFormer proposal has demonstrated a considerable level of efficacy in the classification of PAP smear whole slide images. Unlike other inferior transformer and convolutional models, this particular model has exhibited commendable performance on both private six-class and public four-class datasets. Additionally, the program has showcased strong binary, three-class, and five-class cellular classification precision, recall, accuracy, and F1-scores. The CervixFormer proposal has utilized data augmentation and stain normalization techniques to enhance diversity and staining invariance across the datasets. The incorporation of Swin Transformer subnetworks into the model has facilitated multi-scale feature learning through the fuzzy rank fusion approach. Consequently, the GradCAM visualizations on the important regions have provided clinical interpretability of the model’s outputs. Overall, the CervixFormer proposal has shown promise as a scalable and reliable solution for cervical screening and diagnosis, with the potential for clinical deployment [26]. In brief, the available data suggest that AI exhibits remarkable detection rates and precision when it comes to cytology. Nonetheless, there remains the potential for conducting extensive research and delving into innovative applications within this field. For instance, the development of AI microscopes can potentially revolutionize cytology screening by enhancing its efficiency and accuracy. Additionally, an AI assistant colposcopy represents a highly advanced instrument that can assist in the identification and evaluation of cervical cancer [27]. This particular innovation possesses the capacity to transform the realm of cervical cancer treatment, as it has the ability to furnish healthcare practitioners with invaluable perspectives and assistance. Various potential paths regarding the treatment of cervical cancer exist, highlighting immunotherapy, targeted therapy, and the use of PARP inhibitors [28].