Progress in the application research of cervical cancer screening developed by artificial intelligence in large populations

Abstract

Cervical cancer stands out among various cancers due to its potential for prevention and eradication, mainly through vaccination and proactive screening measures. However, there are still a large number of women in low- and middle-income countries who need to undergo cervical cancer screening. Conventional cervical cancer screening approaches possess distinct benefits and drawbacks regarding sensitivity, specificity, promptness, and expense. In recent years, artificial intelligence (AI) has gained widespread use to help healthcare professionals in performing extensive cervical cancer screenings, focusing on machine learning (ML) and deep learning (DL) techniques. Traditional screening methods combined with AI technology have shown initial effectiveness in cervical cancer screening. But it is necessary to address various challenges such as limited technology and resources, difficulties in integrating clinical workflows, and ethical and legal risks in large-scale population cervical cancer screening. In this review, how AI helps simplify workflows, aids in cytological segmentation and diagnosis, enhances the triage and diagnosis processes for human papillomavirus (HPV) and colposcopy were described firstly. Then we summarized the existing clinical cases of AI applied to large-scale cervical cancer screening. Finally, we discussed the challenges and limitations of AI implementation in large population cervical cancer screening. These insights may possess the capacity to transform cervical cancer screening on a global scale by improving diagnostic precision, facilitating early intervention, and increasing the overall efficacy of cervical cancer screening initiatives worldwide.

Introduction

Cervical cancer ranks among the most prevalent cancers affecting women [1]. Early-stage cervical cancer is linked to an impressive high five-year survival rate of 90%. However, advanced cervical cancer is associated with a dismal five-year survival rate of 16.5%, highlighting the importance of proactive health measures in combating this disease [1, 2]. In recent years, the introduction of HPV vaccination programs and the implementation of effective screening strategies have made cervical cancer become the cancer that can be targeted for elimination [3]. On May 2018, the World Health Organization (WHO) called for coordinated global action to eliminate cervical cancer [4]. But countries around the world are not consistent in achieving this goal. For developed countries, it is expected to be achieved the goal in the coming decades. For most low- and middle-income countries, less than 5% of women undergo screening, these countries face a significant number of preventable cases and fatalities from cervical cancer. It remains unclear whether they will be able to achieve this goal by the end of this century [5]. Therefore, it is urgent to promote large-scale cervical cancer screening in these countries [6].

Nowadays, cervical cancer screening employs several clinical methods to identify precancerous lesions and invasive cancers early. The most commonly employed techniques include the cytology (traditional Pap smear and liquid-based cytology smear), HPV testing and colposcopy [7]. Each screening strategy presents advantages and disadvantages that must be balanced against healthcare resources, patient preferences, and clinical guidelines [8]. In recent years, the application of AI in the field of medicine has become increasingly widespread, which can be used to recognize complex clinical data and images, transforming them into quantifiable and interpretable clinical decisions through algorithms, thereby optimizing and simplifying the workflow of clinicians [9, 10]. In gynecologic oncology, AI has been increasingly applied to enhance the accuracy and efficiency of diagnostic processes, especially in the area of cervical cancer screening [11]. Nevertheless, while currently AI offers significant promise and its potential to revolutionize screening processes, the application of AI in large-scale population cervical cancer screening is still in its infancy [4]. Addressing the challenges and advancing the use of AI in widespread screening efforts are vital for enhancing early detection and ultimately reducing the incidence and mortality associated with cervical cancer [12].

Previous research has focused on summarizing the main systems or algorithms of AI in cervical cancer screening, without delving into how AI can be integrated with traditional cervical cancer screening methods to participate in various aspects of cervical cancer screening work. Additionally, it has not paid attention to the relevant research and importance of AI in large-scale population cervical cancer screening. Therefore, in this review, we comprehensively searched the main literature on the application of AI in cervical cancer screening over the past five years, using the keywords “artificial intelligence”, “cervical cancer” and “screen” in databases such as PubMed and Web of Science. We summarized the new systems, algorithms, and the current status of large-scale clinical applications of AI in cervical cancer screening, discussed the future prospects and directions in this field, which might provide a new perspective for the early clinical application of AI-assisted cervical cancer screening.

Traditional cervical cancer screening methods

Figure 1 presents an analysis of the benefits and drawbacks associated with the existing methodologies for cervical cancer screening. Pap smear is the earliest method for cervical cancer cytology screening, the collecting cells from the cervix were spread onto a microscope slide, stained, and examined for any cellular abnormalities [13, 14]. The specificity of the Pap smear is 75%, whereas its sensitivity is 55.5% [15]. Attributed to sampling errors, differences among observers or laboratory inaccuracies, the potential for false results in Pap smear cannot be overlooked despite its relative simplicity and cost-effectiveness [16]. Liquid-based cytology (LBC) is a modern approach to cervical cancer screening that improves upon traditional Pap smear techniques [17]. The cell samples were placed into a vial containing a liquid preservative, which minimizes contaminants from blood and inflammatory cells, leading to more accurate diagnoses and high specificity of 93.60% [18–20]. However, the limitations of LBC encompass reduced sensitivity, elevated expenses, and the requirement for specialized laboratory equipment and procedures [20]. These factors may restrict the widespread use of LBC.

Fig. 1.

The benefits and drawbacks of current methods in cervical cancer screening

HPV testing is a pivotal method in cervical cancer screening, providing a direct assessment of the virus responsible for most cervical cancer cases [21–23]. HPV testing offers a high sensitivity of 93.85% and it can also provide a longer interval between screenings when the results are negative [23]. However, HPV testing includes a low of specificity 36.09%, it cannot determine the presence of cervical cancer on its own, and may lead to overdiagnosis and overtreatment in cases of transient HPV infections [24].

Colposcopy is used to closely examine the cervix, vagina, and vulva for signs of disease, particularly in women who have received abnormal cytology results or positive HPV tests [25]. If any suspicious lesions are identified during colposcopy, targeted biopsies can be performed for further pathological analysis [26]. However, it relies on the clinician’s expertise and may be subject to variability in interpretation. Patients may experience discomfort during biopsies, and there can be a risk of complications such as bleeding or infection [27, 28].

Basic concepts used in AI for cervical cancer screening

The conception of AI was proposed in 1956 firstly, which was defined as the simulation of human intelligence processes by computer systems [11, 29]. ML constitutes the fundamental component of AI. It enhances the efficacy of computer programs by leveraging data or prior experiences, thereby eliminating the necessity for direct programming by human operators [30, 31]. General classes of ML methods include supervised learning, unsupervised learning, ensemble learning, deep learning and Bayesian learning [32]. ML can achieve good results even with small amounts of data [33]. DL is an emerging special ML method that has developed in recent years, and is the most widely used tool in AI. DL uses neural networks as its architecture to perform representation learning on data [34]. The DL models are more flexible compared to traditional ML models. However, DL models generally necessitate a considerably larger volume of training data [33]. Overall, the ML and DL analysis process can be divided into the following four steps: data input, feature extraction, classification and data output, all without or little human intervention [35]. Figure 2 introduces the algorithms, processes, and correlations of AI, ML and DL. Consequently, selecting the appropriate AI model in alignment with prevalent clinical applications represents a crucial initial step.

Fig. 2.

Relationship and basic process of AI, ML and DL

AI applications in cervical cancer cytology screening

Cell segmentation in cervical cytology images

The overlapping of sampled cervical cells can affect the accuracy of examinations, making it crucial to separate the cytoplasm and nucleus from the overlapping cells. Traditional manual segmentation methods are inefficient and inaccurate, while AI-based cell segmentation methods offer high precision and low error rates, making them particularly suitable for large sample screenings. Therefore, it is essential to utilize AI technology for preprocessing images of overlapping cervical cancer cells, extracting key features to simplify workflows and assist in diagnosis [36]. Zhang et al. developed a DL model based on Mask RCNN and PointRend modules, which could segment overlapping cells and assist pathologists in making quick diagnoses [37]. And using the CNN-based SPCNet model, Zhao et al. achieved the complete segmentation of Papanicolaou smear cells [38]. However, DL-based cell segmentation methods require a large amount of labeled data. The HVS-Unsup model proposed by Yang et al. was an unsupervised cervical cell instance segmentation method based on human visual simulation, which could significantly improve segmentation accuracy and effectively reduce the amount of data labeling needed [39]. Meanwhile, a CNN-based cell segmentation method may lead to false positive results. To optimize the current strategy, Liang et al. constructed a global context-aware framework that could integrate global context information, eliminate false positive samples, and improve the accuracy and specificity of cell segmentation [40].

The insufficient intensity of the overlapping cytoplasmic boundaries is the most challenging aspect of cell segmentation. Therefore, Song et al. proposed the constrained multi-shape evolution algorithm, which could segment overlapping cells based on local and global information priors, significantly reducing erroneous segmentation results and improving cell segmentation performance [41]. Similarly, an algorithm employing U-Net and U-Net + + model architectures with ResNet34 and DenseNet121 has been developed for the automatic segmentation of cervical cell cytoplasm and nuclei utilizing deep ensemble learning techniques. This approach significantly enhanced the process of nuclear-cytoplasmic separation [42].

The large computational load and long processing time make it difficult to apply cell segmentation models in clinical settings, so efforts are being made to improve segmentation efficiency. Zhao et al. have developed a lightweight feature attention network (LFANet) that could accurately segment the nuclear and cytoplasmic regions in cervical images, which achieved state-of-the-art performance with lower computational complexity. This provided a simple and efficient automatic cell extraction method, offering advanced technical support for AI in cervical cancer screening [43]. Similarly, Luo et al. proposed a dual-supervised sampling network structure (DSSNet) that coud significantly reduce the computational load during image segmentation while ensuring the accuracy of the segmentation [44]. Table 1 presents the latest findings regarding AI algorithms and systems utilized in the field of cell segmentation in cervical cytology.

Table 1.

Cell segmentation in cervical cytology images

References	Years	Algorithm or system	Results
Zhang et al. [37]	2024	Novel convolutional neural network based on PointRend	DSC 0.97 TPRp 0.96 FPRP 0.007 FNRo 0.006
Yang et al. [39]	2024	HVS-Unsup	AJI 53.59–73.34% Recall 87.97–92.90%
Song et al. [41]	2024	Constrained multi-shape evolution algorithm	TPR 0.94 ± 0.06 TNR 0.99 ± 0.01 FPR0.01 ± 0.01 FNR 0.28 ± 0.13
Ji et al. [42]	2023	U-Net and U-Net + + with resnet34 and denseNet121 encoders	Cytoplasm segmentation (dice similarity coefficient 0.9535, sensitivity 0.9621, specificity 0.9835) Nucleus segmentation (dice similarity coefficient 0.7863, sensitivity 0.9581, specificity 0.9961)
Zhang et al. [45]	2022	GC-UNet	True Positive 99.1 False Positive 0.5 Precision Rate 99.5
Zhao et al. [38]	2022	SPCNet	DC 93.67% TPp 90.68% FPp 0.18% FN 6.33% AP 90.09%
Zhao et al. [43]	2022	LFANet model	Accuracy 94.85%
Luo et al. [44]	2022	DSSNet	Improve 5 times segmentation speed
Liang et al. [40]	2021	Global context-aware framework	Accuracy improves 5.7% Specificity improves 18.5%
Hoque et al. [46]	2021	Contour property-based approach for segmentation of nuclei	Precision 94.9% Recall 75.9% Aggregated Jaccard Index 68.1%

DSC dice similarity coefficient; TPRp pixelwise true positive rate; FPRP pixelwise false positive rate; FNRo object false negative rate; DSSNet dual-supervised sampling network structure; DC dice coefficient; TPp true positive rate; FPp false positive rate; FN false negative rate

The development of AI-assisted cytology microscopy

It is well known that the microscope is the most important tool for cervical cancer cytology screening. After obtaining the cervical smear, the first step is to conduct a preliminary observation under low magnification microscope, followed by a focused examination of suspicious areas under high magnification [47]. Using CNN based on DL, Kurita et al. could accurately analyze normal and abnormal images at low magnification, allowing for the separation of negative specimens, the accuracy of the model could reach 87.3% [48]. And Tang et al. developed an AI microscope equipped with augmented reality (AR) technology, which projected the results identified by the AI system into the microscope, assisting pathologists in observing and diagnosing cytological results [49]. These gave cytologists enough time to observe abnormal specimens under high magnification, thereby assisting doctors in saving time and simplifying the workflow. Table 2 presents the latest findings regarding algorithms and systems utilized in the field of AI-assisted cytology microscopy.

Table 2.

AI-assisted microscopy in cervical screening

References	Years	Algorithm or system	Results
Kurita et al. [48]	2023	Noisy student training technique based on DL	AUC 0.908 Accuracy 0.873
Tang et al. [49]	2021	AI microscope	Accuracy 0.923 & 0.919

AUC area under curve

Combination other detection techniques with AI in cytology screening

Currently, in addition to using Pap smears or LBC for cervical cancer cytological screening, there are still many other advanced screening technologies available. Firstly, raman spectroscopy is a simple and non-invasive method for early cancer screening. Different types of cervical cancer tissues can be classified based on their nucleic acids and proteins using Raman spectroscopy [50]. By linking Raman spectroscopy with DL and constructing an efficient channel attention convolutional neural network (ECACNN) model, it could be used to screen and classify cervical precancerous lesions and early cervical cancer, which was a feasible and high accuracies of 94.04% method [51]. Similarly, Yan et al. also used Raman spectroscopy to collect cervical tissue slice data by constructing a one-dimensional hierarchical convolutional neural network (H-CNN) in order to classify and identify the tissue samples [52].

The high-resolution endomicroscope (HRME) is capable of imaging the nuclear morphology associated with cervical lesions. By combining HRME with DL, which could identify intraepithelial neoplasia grade 2 or more severe with sensitivity and specificity of 94% and 58%. Additionally, combining diagnostic data with HPV test results could significantly improve the specificity to 71% of detecting cervical precancerous lesions [53]. The Fluorescence Lifetime Imaging Microscopy (FLIM) technology was combined with ML techniques to construct the FLIM-ML Model. This model evaluated the fluorescence intensity and metabolic state of unstained exfoliated cervical cells through FLIM, and then ML analyzed the FLIM images. It could predict the risk of cervical cancer occurrence with a sensitivity and specificity of 90.9% and 100%, respectively. The FLIM-ML Model might have broad clinical application prospects in predicting the occurrence, recurrence, and diagnosis of cervical cancer in the future [54]. Moreover, modern multimodal hyperspectroscopy (MHS) is an AI-based technology, the MHS inspection device identified lesions through the detection of metabolic and morphological indicators within cervical tissue, which significantly reduced the false negative rate in cervical screening [55]. Table 3 presents the latest findings regarding techniques utilized in the field of the combination other detection techniques with AI in cytology screening.

Table 3.

The combination of other detection techniques with AI in cervical screening

References	Years	Algorithm or system	Results
Kang et al. [51]	2023	ECACNN	Accuracy 94.04% AUC 96.86%
Yan et al. [52]	2024	H-CNN	Accuracy 96.49%
Brenes et al. [53]	2022	DL-HRME model	Sensitivity 94% Specificity 71%
Ji et al. [54]	2022	FLIM-ML model	Sensitivity 90.9% Specificity 100%
Krasznai et al. [55]	2021	MHS	False negative rate 1.80%

ECACNN efficient channel attention convolutional neural network; H-CNN hierarchical convolutional neural network; HRME high-resolution endomicroscope; FLIM fluorescence lifetime imaging microscopy; AI-DS artificial intelligence-dual stained; MHS modern multimodal hyperspectroscopy

AI-assisted triage of cytological negative cases

Many countries around the world are facing the issue of pathologists having heavy workloads and varying levels of experience. As a preliminary screening tool, AI can classify cases and provide initial diagnoses, which may greatly alleviate this dilemma [56]. Bai et al. developed an Artificial Intelligence Assistive Cytology Diagnostic System (AICyte), which was based on the AI deep learning CNN Yolo5 model to detect and classify the suspicious cells. It provided 20 AI diagnostic images for manual review. AICyte had the sensitivity of 97.89% and specificity of 16.19% higher than that of manual screening. A cytologist-in-the-loop artificial intelligence (CITL-AI) system was showed by Xue et al., which could initially screen negative cases, achieve case triage, and alleviate the workload of pathologists. At the same time, for less experienced cytologists, CITL-AI could assist in diagnosis to improve sensitivity and specificity, which allowed for maintaining high-quality cervical cancer screening results remotely in areas with relatively underdeveloped healthcare conditions [58]. Similarly, the L-PCNN algorithm developed by Li et al. could achieve the goal of cervical cell automatic classification and had a 98.89% effectiveness of accuracy [59]. The attention feature pyramid network (AttFPN) built by Cao et al. also could automatically detect abnormal cervical cells in cytological examinations to assist pathologists in making quick diagnoses [60]. Table 4 presents the latest findings regarding algorithms and systems utilized in the field of AI-assisted triage of cytological negative cases.

Table 4.

AI-assisted cytology triage system

References	Years	Algorithm or system	Results
Bai et al. [57]	2024	AICyte system	Sensitivity 97.89% Specificity improves 16.19% Accuracy improves 7.09%
Xue et al. [58]	2024	CITL-AI system	Sensitivity 81.6% Specificity 78.9%
Li et al. [59]	2021	L-PCNN algorithm	Accuracy 98.89% Sensitivity 99.9% Specificity 99.8%
Cao et al. [60]	2021	AttFPN	Accuracy 90.91% Sensitivity 91.30% Specificity 90.62%

AICyte artificial intelligence assistive cytology diagnostic system; CITL-AI cytologist-in-the-loop artificial intelligence; AttFPN attention feature pyramid network

Cervical cancer cytology classification based on transfer learning

Transfer learning is a special type of ML. Traditional machine learning requires a large amount of data to make accurate predictions, while transfer learning has become an innovative method to alleviate data scarcity [61]. The Swin Transformer, as an excellent computer vision model of transfer learning, combined with genetic algorithm (GA) and random forest classifier, forms the Swin-GA-RF model, which could accurately classify cervical cancer cytology results and had a very broad clinical application prospect [62]. Kurita et al. have adopted a visual language model that can perform high-precision and high-speed evaluations of cytological images. It ranks abnormal cases based on the characteristics of individual cases to simplify the cervical cancer screening process, providing possibilities for expanding early detection and treatment of cervical cancer [63]. Allogmani et al. preprocessed the images utilizing Archimedes optimization algorithm with transfer learning (CPLDC-AOATL) methodology to preprocess, extract features, and classify cytological screening data, which could more effectively screen and identify precursors to cervical cancer [64]. Table 5 presents the latest findings regarding algorithms and systems utilized in the field of transfer learning.

Table 5.

The application of transfer learning in cervical cancer screening

References	Years	Algorithm or system	Results
Alohali et al. [62]	2024	Swin-GA-RF	Accuracy 99.012 Precision 99.015 Recall 99.012 F1-score 99.011
Kurita et al. [63]	2024	Visual language models	10 s per case
Allogmani et al. [64]	2024	CPLDC-AOATL algorithm	Accuracy 99.53%

Swin-GA-RF swin transformer- genetic algorithm- randomforest; CPLDC-AOATL archimedes optimization algorithm with transfer learning

The combination of more than one algorithm of DL and ML

DL demonstrates a remarkable capability in the extraction of intricate features from images, thereby establishing a foundational basis for a range of ML algorithms. And combining DL and ML, Mathivanan et al. innovatively used the SIPaKMeD dataset to first extract key features from images through DL using CNN architectures. Then, based on ML algorithms accurate diagnosis and classification of cervical cancer could be achieved. The DL-ML method had an impressive 98% accuracy rate and has broad clinical application prospects in advancing cervical cancer screening in large populations [65]. Each algorithm has its superiorities and weaknesses, the combination and complementarity of different algorithms may be an important way to improve the performance of cervical cancer screening. Combination of YOLOv4 and ResNeSt two DL algorithms, Nambu et al. first diagnosed atypical cells based on the YOLOv4 algorithm, and then used the ResNeSt algorithm to classify the aforementioned cells. The two-step strategy could accurately classify cervical liquid-based cytology results with an accuracy of 90.5% [66]. Table 6 presents the latest findings regarding algorithms and systems utilized in the combination of more than one algorithm of DL and ML.

Table 6.

The combination of DL and ML algorithms assists in cervical cancer screening

References	Years	Algorithm or system	Results
Mathivanan et al. [65]	2024	DL-ML	Accuracy 98.08%
Nambu et al. [66]	2021	YOLOv4 and ResNeSt	Accuracy 90.5%

DL-ML machine learning-deep learning

AI applications in HPV testing

The role of AI in HPV detection could be enhanced by highlighting its specificity in improving diagnostic accuracy beyond traditional methods. The triage of HPV-positive patients is an important component of cervical cancer screening [67]. In the work of Ahmed et al., an automated visual evaluation (AVE) system was utilized, applying concentrated multi-level, ordinal ground truth delineation schemes in the model selection. By combining HPV typing and risk stratification with predictive models, a reliable, reproducible, and clinically translatable cervical cancer screening model was developed, serving as an effective triage tool for HPV-positive women [68]. However, HPV infection is often transient, and only a very small portion will progress to malignancy [69]. Using cytological examinations to triage HPV-positive patients could significantly reduce the excessive examinations and medical burden associated with colposcopy [70]. Xue et al. trained AI using data from HPV-positive patients with abnormal cytology results to build an artificial intelligence-enabled liquid‐based cytology (AI‐LBC) system that could triage HPV-positive patients with either normal or abnormal cytology. Comparing the AI‐LBC, human cytologists, and HPV16/18 typing for HPV-positive patients, it was found that AI-LBC has comparable sensitivity of 86.49%, and also reduced referrals to colposcopy by approximately 10%. This might play an important role in assisting cytologists with the triage of HPV-positive patients [71]. Therefore, AI can help cytologists make more accurate triage decisions for HPV-positive patients by analyzing complex data patterns and combining cytological and molecular information, thereby reducing false positives and unnecessary colposcopies, and improving diagnostic specificity. Table 7 presents the latest findings regarding algorithms and systems utilized in AI applications in HPV testing.

Table 7.

HPV testing-assisted AI system or algorithm

References	Years	Algorithm or system	Results
Ahmed et al. [68]	2023	AVE system	AUC 0.89
Xue et al. [71]	2023	AI-LBC	Sensitivity 86.49% Reduced referrals to colposcopy by approximately 10%

AVE automated visual evaluation; AI-LBC artificial intelligence‐enabled liquid‐based cytology; AUC area under curve

AI applications in colposcopy

The results of colposcopy largely depend on the clinical experience of the colposcopist. However, a large number of patients and doctors with varying levels of medical expertise cannot meet clinical needs in a timely manner [72]. Ouh et al. developed an AI-based analysis software called CerviCARE AI, which could automatically analyze cervicography images, classify cervical intraepithelial lesions, has significant clinical application value and utility with a remarkable sensitivity of 98% and a specificity of 95.5% [73]. An automated visual evaluation (AVE) screening method was created, after taking cervical images with a mobile phone and uploading them to the AVE system, screening results could be obtained quickly. Although the system still had certain limitations, such as the sensitivity of 85% and specificity 86%, it could assist in diagnosing results from visual inspection with acetic acid, and it might have great development potential in the future [74]. Chandran et al. have also developed a model called Colposcopy Ensemble Network (CYENET), which obtained images from colposcopy and classifies cervical cancer based on Ensemble DL, achieving an accuracy of 92.3%, a sensitivity of 92.4%, and a specificity of 96.2% [75]. And Xue et al. developed a Colposcopic Artificial Intelligence Auxiliary Diagnostic System (CAIADS), which could extract and analyze feature information from colposcopic images based on DL, providing diagnostic information and images of targeted biopsy areas to assist less experienced physicians in performing colposcopy and biopsies. The system had been validated on a large dataset and demonstrated excellent performance [76]. Similarly, the AI system developed by Takahashi et al. was capable of autonomously identifying high-grade lesion areas across 12 segments of the cervicovaginal region, achieving a detection rate of 62.1% and an overall accuracy rate of 89.7%. This system demonstrated significant potential in aiding the identification of patients with cervical intraepithelial neoplasia (CIN) or more advanced lesions during colposcopic examinations [77].

The transformation zone of the cervix is a common site for cervical malignancy, and its type is an important factor in the grading of colposcopic examination results and the selection of treatment [78]. Cao et al. have developed a classification network based on DL that could automatically identify the types of cervical transformation zones based on original colposcopic images. The method reached 88.49% accuracy and the sensitivities for type 1, 2 and 3 of cervical transformation zones were recorded at 90.12%, 85.95%, and 89.45%, which played an important role in assisting cervical colposcopy screening and diagnosis [79]. What is currently clear is that we cannot fully rely on independent AI models based on colposcopy to accurately screen for cervical cancer. Therefore, Fu et al. used cervical physiological saline, iodine, and acetic acid images to construct a multivariable logistic regression (MLR) model. Then, by combining cytology and HPV tests, they developed a cross-modal integrated model using MLR based on DL to improve the performance of cervical cancer screening. The AUC of the model is 0.921, which could achieve precise cervical cancer screening [80]. Table 8 presents the latest findings regarding algorithms and systems utilized in AI applications in colposcopy.

Table 8.

Colposcopy-assisted AI system or algorithm

References	Years	Algorithm or system	Results
Ouh et al. [73]	2024	CerviCARE AI	Sensitivity 98%, specificity 95.5%
Hu et al. [74]	2021	AVE	Sensitivity 85%, specificity of 86%
Chandran et al. [75]	2021	CYENET	Accuracy 92.3%, specificity 96.2%, kappa scores 88%
Xue et al. [76]	2020	CAIADS	Sensitivity low-grade or worse 90.5%, high-grade or worse 71.9% Specificity low-grade or worse 51.8%, high-grade or worse 93.9%
Takahashi et al. [77]	2022	Classification network based on DL	Detection rate 62.1% Accuracy rate 89.7%
Cao et al. [79]	2023	A deep learning-based method	Accuracy 88.49% Sensitivities for type 1, 2, 3 90.12%, 85.95%, 89.45%
Fu et al. [80]	2022	MLR	AUC 0.921

CAIADS colposcopic artificial intelligence auxiliary diagnostic system; AVE automated visual evaluation; CYENET colposcopy ensemble network; CAIADS colposcopic artificial intelligence auxiliary diagnostic system; MLR multivariable logistic regression; AUC area under curve

Clinical application cases of artificial intelligence in large-scale cervical cancer screening

The aforementioned cervical cancer screening systems are based on AI algorithms and models constructed from retrospective clinical samples, and then validated using a small number of clinical samples. However, the prospective collection of clinical samples and the application of the developed system in clinical settings is where the clinical significance of cervical cancer screening lies. Therefore, we have listed some current cases of applying AI to large-scale population screening for cervical cancer, in hopes of providing better clinical guidance for other countries.

From 2017 to 2021, a large-scale of 1,474,788 women in Hubei Province underwent cervical cancer screening under AI-assisted cytology diagnosis. After professional healthcare workers collected and prepared cervical cytology slides, diagnosticians used AI to screen for cytologically positive patients. Then, cytologists reviewed positive cases and randomly checked negative cases for quality control. Each slide was scanned within 2 min and diagnosed within 1.5 min, and the screening results could be promptly feedback to patients online or via text message. This study confirmed that AI-assisted cytological screening was efficient, cost-effective of 6.31$ per woman, and highly accurate, which would significantly benefit women in poor, remote, and medically underserved areas. It might have an optimistic outlook for future application across China or even worldwide [ 81 ].

Similarly, Bao et al. conducted a study involving the collection of cervical cytology samples from 701,301 women in Hubei Province, which took place between 2017 and 2018, with the aim of implementing AI-assisted cytology screening. The findings of this research revealed that the detection rate of CIN2 + was 20% higher in ASC-H grade and 40% higher in HSIL grade when assessed by AI compared to traditional manual interpretation. This large-scale population study preliminarily confirmed that AI could assist in improving the diagnostic rate of cervical cancer screening [82]. Concurrently, this study assessed the cost-effectiveness of AI in facilitating cervical cancer screening. The findings suggested that AI-assisted LBC screening conducted every five years may prove to be more cost-effective than traditional manual interpretation of LBC results [83].

Zhu et al. had independently developed an Artificial Intelligence-Assistive The Bethesda system (AIATBS) system and used this system to train on 81,727 retrospective samples of cervical liquid-based cytology to extract high-quality data features. They then collected 34,403 prospective samples from 11 medical institutions for large-scale cervical cancer screening and clinical validation. The research results indicated that the AIATBS system has extremely high sensitivity (intraepithelial lesions ≥ 92.00%) and specificity of 82.14%, accurately classifying cervical cytology results into TBS categories, comparable to the diagnostic capabilities of cytopathologists. This study demonstrated the potential for AI-assisted large-scale cervical cancer diagnosis [84].

Wentzensen et al. developed the CYTOREADER cloud platform, which utilized DL techniques to facilitate the automatic analysis of p16/Ki-67 dual-stained (DS) cytology slides through whole slide scanning and deep neural networks. Validation conducted on a sample size of 4,253 samples showed that the automated DS system significantly improved specificity while maintaining sensitivity, achieving a specificity of 61.5% in the HPV screening cohort. Furthermore, the referral rate for colposcopy decreased from 60.1 to 41.9%. The system demonstrated stable performance across different liquid-based technologies and simulated HPV vaccination populations. The study indicated that by eliminating subjective human judgment, this system could effectively reduce unnecessary invasive examinations, providing an efficient solution for global cervical cancer screening [85].

Research team at Sun Yat-sen Memorial Hospital in China, has recently created an artificial intelligence cervical cancer screening system (AICCS) designed to facilitate the diagnosing grading of cervical cancer cytology. In the model training and validation conducted on retrospective, prospective, and randomized observational trial datasets in 16,056 eligible participants, it was found that AICCS-assisted cytopathologists demonstrated higher AUC of 0.947, specificity of 89.0%, and accuracy of 89.2%, with comparable sensitivity of 94.6%. These results indicated that the AICCS system is significant in assisting pathologists with accurate diagnosis and grading of cervical cancer screening. AICCS might become an important auxiliary system for large-scale population screening in the future [86].

A pilot study on cervical cancer screening, conducted by Poli et al. in southern India, sought to develop and validate an artificial intelligence-based visual inspection tool (VIA-AI) aimed at minimizing overtreatment in cervical cancer screening. The study involved 2,052 participants and the findings indicated that the clinical accuracy of the VIA-AI tool was 76%, with a sensitivity of 62.5%, specificity of 97.6%, and an area under the ROC curve of 0.76. The study concluded that the VIA-AI tool holds significant potential for application in resource-constrained settings, offering decision support and mitigating unnecessary treatments [87].

In summary, the above cases of AI applications in large-scale cervical cancer screening demonstrated that AI might pave the way for new cervical cancer screening in countries and regions with large rural populations and low economic development, which may provide a new strategy to achieve the global elimination of cervical cancer. Table 9 presents some application cases of AI in large-scale cervical cancer screening.

Table 9.

Clinical application cases of artificial intelligence in large-scale cervical cancer screening

References	Sample size	Algorithm	Results
Zhu et al. [81]	1,474,788	Bethesda system	10–12 s and $6.31 per sample
Bao et al. [82]	701,301	AI cytology system	Sensitivity ≥ 88.90% Specificity ≥ 94.40%
Zhu et al. [84]	34,403	AIATBS	Sensitivity ≥ 92.00% Specificity 82.14%
Wentzensen et al. [85]	4,253	CYTOREADER cloud platform	Specificity 61.5% Referral rate for colposcopy 41.9%
Wang et al. [86].	2780	AICCS	AUC 0.947 Sensitivity 94.6% Specificity 89.0% Accuracy 89.2%
Poli et al. [87]	2,052	VIA-AI	Sensitivity 62.5%, Specificity 97.6% ROC curve 0.76

AIATBS artificial intelligence-assistive the bethesda system; AICCS artificial intelligence cervical cancer screening system; VIA-AI artificial intelligence-based visual inspection tool; ROC receiver operating characteristic

Perspective and challenges of AI applied to large-scale population screening

Generally speaking, large-scale AI cervical cancer screening requires operation on a high-sensitivity computing platform, covering multiple regions or countries, to quickly identify high-risk populations and improve screening efficiency. In contrast, small-scale AI cervical cancer screening focuses more on patients with known precancerous lesions, follow-up populations, or individuals with personalized needs. This requires collaboration between AI technicians and pathologists to achieve a high diagnostic rate. Despite the global initiatives to accelerate the elimination of cervical cancer, less than 5% of women undergo screening in low- and middle-income countries, therefore, it is urgent to expand the scope of cervical cancer screening and to identify high-risk groups at an early stage [5].

According to the summary of our discussion above, AI-assisted large-scale cervical cancer screening has broad application prospects. Firstly, AI can quickly process a large number of cervical cell images and data, significantly improving the efficiency of cervical cancer screening. Secondly, AI can identify subtle pathological features that are difficult for the human eye to detect, which may reduce the occurrence of missed and misdiagnoses, thereby increasing the diagnostic rate. Additionally, in grassroots and low-income countries, the combination of AI with smartphones can alleviate the pressure of resource shortages, which is beneficial for the rational allocation of medical and health resources. Finally, AI-assisted screening and diagnosis may provide a new strategy to achieve the global elimination of cervical cancer.

However, AI screening for cervical cancer still faces many challenges. First, AI screening for cervical cancer is based on large-scale images and data to train models. However, the formats and standards of these images and data have not been unified, and the professional levels of healthcare personnel collecting the data are inconsistent, which may affect the accuracy of the models. Secondly, the current models and algorithms used for cervical cancer screening each have their own advantages and disadvantages, and there is still a need to continue developing universal, general, and efficient algorithms. Thirdly, AI-assisted cervical cancer screening is a process of human-machine interaction, and the way clinical doctors use the machines can also affect the screening results. Finally, the security of patient information, images, accessibility, and data, as well as the ethical issues involved with AI, also require relevant laws and regulations for regulation. Figure 3 shows the limitations and advantages of AI in cervical cancer screening.

Fig. 3.

The advantages and disadvantages of AI in cervical screening

Therefore, AI screening has the potential for long-term cost-effectiveness and accuracy. It can handle complex data, reduce labor costs, and improve diagnostic accuracy, but it requires significant upfront investment and technical support for implementation. Traditional methods have low short-term costs, are technically mature, and are easy to implement, but they incur high labor costs and low efficiency during large-scale screenings. By combining the advantages of both approaches, the effectiveness of cervical cancer screening may be enhanced. However, our article also has its shortcomings. We need to further explore the usage of various AI models for cervical cancer screening in order to provide references for clinical selection. At the same time, we need to gain a deeper understanding of the clinical translation methods of AI models for cervical cancer screening.

Conclusion and directions

In summary, based on the current clinical examples of AI applications in cervical cancer screening and the increasingly mature AI algorithms, we have reason to believe that the use of AI for large-scale cervical cancer screening is just around the corner. Future research on AI in cervical cancer screening may need to focus on the following directions: the development of more efficient AI algorithms aimed at enhancing diagnostic accuracy; the establishment of extensive, multi-center high-quality datasets to bolster the generalizability of predictive models; the optimization of AI system integration within clinical workflows to augment their practicality and acceptance; the execution of multi-center clinical trials to assess the long-term impacts and cost-effectiveness of AI-driven screening methods; and the investigation of the synergistic potential of AI with other technological innovations to create a more intelligent and automated screening framework.

Author contributions

WXL: conceptualization, data curation, methodology, writing– original draft; XYX: conceptualization, project administration.

Funding

This work was supported by the Key Research and Development Program of Hubei Province (2022BCA041).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Human ethics and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing Interests

The authors declare no competing interests.