Cancer Biol Med 2024. doi: 10.20892/j.issn.2095-3941.2024.

0198

REVIEW

Artificial intelligence strengthens cervical cancer screening –

present and future
Tong Wu1, Eric Lucas2, Fanghui Zhao3, Partha Basu2, Youlin Qiao1
1School of Population Medicine and Public Health, Chinese Academy of Medical Sciences and Peking Union Medical College,
Beijing 100730, China; 2Early Detection, Prevention & Infections Branch International Agency for Research on Cancer (WHO),
25 avenue Tony Garnier, Lyon 69007, France; 3Department of Cancer Epidemiology, National Cancer Center/National Clinical
Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College,
Beijing 100021, China

ABSTRACT Cervical cancer is a severe threat to women’s health. The majority of cervical cancer cases occur in developing countries. The
WHO has proposed screening 70% of women with high-performance tests between 35 and 45 years of age by 2030 to accelerate the
elimination of cervical cancer. Due to an inadequate health infrastructure and organized screening strategy, most low- and middle-
income countries are still far from achieving this goal. As part of the efforts to increase performance of cervical cancer screening,
it is necessary to investigate the most accurate, efficient, and effective methods and strategies. Artificial intelligence (AI) is rapidly
expanding its application in cancer screening and diagnosis and deep learning algorithms have offered human-like interpretation
capabilities on various medical images. AI will soon have a more significant role in improving the implementation of cervical cancer
screening, management, and follow-up. This review aims to report the state of AI with respect to cervical cancer screening. We
discuss the primary AI applications and development of AI technology for image recognition applied to detection of abnormal
cytology and cervical neoplastic diseases, as well as the challenges that we anticipate in the future.
KEYWORDS Cervical cancer screening; artificial intelligence; deep learning algorithms

using high-performance testing, early diagnosis, and timely

Introduction treatment of high-grade cervical intraepithelial neoplasia
(CIN 2/3) and cancer. Cervical cytology has been widely
Cervical cancer remains a major cause of mortality among
used for cervical cancer screening but global implementa-
women with > 662,000 new cases diagnosed and approxi-
tion poses several challenges and suboptimal sensitivity of
mately 349,000 deaths reported globally in 20221. The long
the test necessitates frequent screening. The WHO has rec-
premalignant phase of cervical cancer and the natural pro-
ommended HPV detection-based tests for primary screening
gression of the disease make cervical cancer the only cancer
due to the higher sensitivity and objective nature of the test.
that is currently preventable through primary and second-
While cytology triage of HPV-positive women to colposcopy
ary prevention2. The World Health Organization (WHO)
before confirmation of abnormal cervical cytology serves
has launched a global strategy to eliminate cervical cancer
as an effective primary approach in high-income countries,
by the end of the millennium that includes vaccination
deploying this strategy requires a well-organized infrastruc-
against human papillomavirus (HPV), screening of women
ture and the expertise of professionals, including patholo-
gists, cytopathologists, laboratory scientists, and experienced
Correspondence to: Youlin Qiao and Partha Basu
colposcopists3. The WHO also recommends visual inspec-
E-mail: qiaoy@cicams.ac.cn and basup@iarc.who.int
ORCID ID: https://orcid.org/0000-0001-6380-0871 tion with acetic acid (VIA) for triaging HPV-positive women
and https://orcid.org/0000-0003-0124-4050 because VIA is a cost-effective approach suitable for lim-
Received May 30, 2024; accepted August 12, 2024; ited resourced settings. However, VIA has strong diagnos-
published online September 19, 2024.
tic subjectivity and may lack precision, potentially resulting
Available at www.cancerbiomed.org
©2024 The Authors. Creative Commons Attribution-NonCommercial 4.0 in pre-cancerous lesions going undetected for an extended
International License period of time4. Colposcopy, as a diagnostic tool, also has
Cancer Biol Med Vol 21, No 10 October 2024 865

the limitations of a subjective test requiring high level of at the population level. It is essential that AI should align with
competency. the standards required for adjunctive use in routine clinical
Recent developments in artificial intelligence (AI) offer settings. The precision of AI in screening tasks, particularly
considerable prospects for an automated, objective, and unbi- regarding misdiagnoses and misdiagnosed cases, is a critical
ased detection of cervical cancer and precancerous conditions. issue in future enhancements. Also, the willingness and trust
The idea of computers simulating human behavior, cognition, of clinicians to embrace AI reflects the hurdles that this novel
and actual thinking was proposed by Alan Turing as early as technology must overcome to be fully applied to cervical can-
1950. The term, “artificial intelligence” was officially coined by cer screening solutions. In this review we reviewed the current
John McCarthy at an academic conference in 19565. AI began state of research and the application of AI in cervical cancer
with the following major directions: perceptrons; Bayesian screening, analyzed the ongoing challenges related to tech-
networks; pattern recognition; human-computer interaction; nological advancement, and advocate for the promotion and
knowledge representation; and computer vision. As AI entered acceleration of widespread use of AI to screen and diagnose
the “golden” era, there was a surge in interest and the perfor- cervical cancer.
mance of AI gradually evolved into complex algorithms that
resemble the logic of human beings. With the development of
machine learning as a core technology of AI, computers can
Machine learning in cervical cancer
learn from data analysis, derive standards from the data, and risk prediction
use these standards to predict and classify unknown objects.
Computers have the capability of finding features and learn- Cervical cancer screening typically involves a series of pro-
ing and distinguishing new text, images, signals, and other cedures, including HPV testing, visual examinations, cytol-
data automatically. From the development of artificial neu- ogy, colposcopy, and biopsies. Each method requires skill
ral networks, the concept of deep learning emerged and is and experience and/or substantial resources and time. In
now widely used in the fields of medical diagnosis, medical resource-limited areas, the implementation of comprehensive
image recognition, natural language processing, and health and high-quality screening programs presents considerable
management applications. In recent years, AI has demon- challenges. Leveraging existing clinical data for efficient and
strated significant advantages in several aspects of detecting intelligent screening or prediction is of substantial value to
cervical cancer, including the segmentation and classification overcome the challenges. The results of HPV testing and HPV
of cytology6,7, ­colposcopy8, and the early detection of cervi- genotyping combined with other clinical information, such as
cal cancer lymph node metastasis (LNM) on magnetic reso- age, menstrual status, and behavior, can be utilized to predict
nance imaging (MRI)9. A significant proportion of current the progression of positive high-risk (hr)HPV cases and the
research focuses on developing deep learning algorithms for risk of cervical cancer13,14. Moreover, prediction of cervical
automatic processing, recognition, feature extraction, and cancer risk by integrating HPV test results with cytologic find-
classification of cervical images, which enables AI to analyze ings and biomarkers has been shown to improve upon con-
images, identify patterns, and interpret cancer characteristics. ventional screening methods, thereby reducing the referral
The WHO has noted that AI can enhance screening tests and rates for colposcopy15,16. A predictive model to identify those
techniques that involve visual evaluation of digital images10. It at high risk of developing cervical cancer has been developed
is anticipated that AI-assisted screening will have a major role based on prior HPV results and historical medical records,
in low-resource areas, addressing the shortage of competent allowing for individualized risk stratification and manage-
healthcare personnel. Additionally, the internet and mobile ment17. These predictive models may guide development of
data, cloud computing, and mobile devices have improved risk-stratified cervical cancer screening strategies.
access to healthcare services in remote areas, thereby reduc- Technically, most predictive models are constructed by
ing healthcare costs, while remote digital education platforms machine learning algorithms, such as support vector machines
can enhance the professionalism of local physicians and alle- (SVMs) and random forests, which form the foundation of
viate the global shortage of specialists11,12. Despite the posi- deep learning and represent the precursors of AI develop-
tive developments in AI for cervical cancer screening, further ment. Machine learning algorithms are relatively interpretable
exploration and validation are needed to prove its effectiveness in medical applications and perform well in classification and
866 Wu et al. Artificial intelligence strengthens cervical cancer screening

prediction tasks. However, intelligent analysis of diverse data unexpected data. Presently, AI in cervical cancer research is
types presents challenges, particularly in the highly abstract primarily focused on the automatic detection, feature extrac-
feature extraction required for unstructured data, which tion, and learning classification of various cervical images.
requires implementation of neural network architectures. To Intelligent analysis of cervical images by advanced computer
intelligently address various tasks in cervical cancer screening, vision techniques is becoming an auxiliary or even alternative
deep learning solutions are more adept at handling different method for detecting cervical cancer at an early stage.
types of unstructured data and integration of multimodal
data. Several studies have combined machine learning and Availability of datasets for AI-guided cervical
deep learning approaches to enhance the robustness of diag- cancer screening
nostic classification tasks, combining deep neural networks
for feature engineering and machine learning algorithms for The availability of large and high-quality datasets of cer-
classification tasks, which result in more accurate and inter- vical clinical data provides a solid foundation for training
pretable classifications18-20. Through the application of these and validating AI algorithms. Several high-quality public
technologies, cervical cancer screening can be more efficient datasets with annotations are available, including the Cx22
and accurate, providing essential support for early detection dataset5 and ISBI Challenge Database6,7 for segmenting
and intervention. cytology images and the SIPaKMeD dataset8 and Harlev
datasets9 for classifying cytology cells based on morphology.
However, datasets for colposcopy images are relatively lim-
AI-guided technologies in cervical ited. Presently, the largest public dataset, Intel & MobileODT
cancer screening Cervical Cancer Screening10, is collected by mobile-level
colposcopy devices. However, public access to datasets cap-
The tests used for cervical cancer screening include HPV test- tured with high-magnification colposcopy equipment is still
ing, cytology (both conventional and liquid-based c­ ytology), lacking. The International Agency for Research on Cancer
and VIA (by naked eye or enhanced with a magnifying device). (IARC) Cervical Cancer Image Bank11 is one such database
Lugol’s iodine can also be used in place of acetic acid (VILI), compiled by collaborating colposcopists using standard for-
although Lugol’s iodine is not widely recommended. Recently, mats, although the scale is quite modest.
cervical cancer screening has become increasingly depend-
ent on the detection of hrHPV, which has a higher sensitiv- Feature representative of AI model
ity and negative predictive value compared to cytology. Most
programs recommend triaging HPV-positive women with Representative variables, such as female age, menopausal
a combination of HPV16/18 testing and cytology followed status, parity, medical history, and HPV results, are typi-
by colposcopy. However, cytology tests in low- and middle-­ cally selected as relevant clinical features for a cervical can-
income countries (LMICs) have highly variable performance cer risk prediction model, ensuring a comprehensive and
and low sensitivity due to lack of trained personnel, infrastruc- robust representation of the data. However, for whole-slide
ture, and quality assurance. In addition to colposcopy and and colposcopy images, deep learning algorithms, like con-
cervical biopsy for diagnostic purposes, a colposcopy-guided volutional neural networks (CNNs), are required to extract
biopsy is critical for determining whether further treatment is high-dimensional features from the images. With increasing
necessary. Therefore, colposcopists need comprehensive train- network depth, the selected features become more represent-
ing to achieve a requisite level of proficiency to perform diag- ative, ultimately capturing unique patterns and textures that
nostics capabilities. Nevertheless, colposcopic equipment and can contribute to the diagnosis of cervical cancer. Several pre-
expert or well-trained colposcopists are both scarce resources processing steps are performed before the images are fed into
for LMICs. the neural network, such as normalizing the images, resizing
AI is the simulation of human-like cognitive and learning the images to a consistent pixel size, and applying augmenta-
capabilities by computer systems. AI refers to the capability tion techniques to enhance the robustness of the model. After
of machines to sift and discern patterns from representa- the features have been extracted, the features are normalized
tive examples to assimilate knowledge features and foresee and combined to form a feature vector, which is used as input
Cancer Biol Med Vol 21, No 10 October 2024 867

in the AI model. By utilizing imaging data appropriately, the cancer screening includes a variety of tests, including HPV
multi-faceted nature of the disease can be captured and per- testing as the primary screening, cytology triage, or HPV and
formance of the AI model can be enhanced. cytology co-testing. Colposcopy is used as a preliminary diag-
nostic and the screening results are also required for reference.
AI algorithms for cervical cancer detection Transformer neural network (Transformer) captures depend-
and diagnosis encies between different positions in sequence data through
self-attention mechanisms, while multi-head attention mech-
With deep learning, various features in images, such as color, anisms allow the model to focus on different parts of the input
texture, and relative objects, are systematically captured by neu- sequence simultaneously. This enables Transformers to effec-
ral networks. The CNN, a leading deep learning architecture, tively handle long-range dependencies, making Transformers
extracts high-level features, such as edges and textures, from highly effective in natural language and image processing tasks.
cervical cell images through multiple layers of convolution and Specifically, Transformers can integrate different modalities of
pooling operations. The CNN is widely utilized for cell detec- data, such as HPV testing results, cytology, and cervical images,
tion, segmentation, classification, and extraction of regions of showcasing the strong capabilities of Transformers in multi-
interest (ROIs) in cytologic images. Moreover, advanced deep modal data processing and complex sequence m ­ odeling12,13.
learning methods, such as graph neural networks (GCNs), As a result of evaluating multiple performance metrics com-
perform convolution operations on graph-structured data, prehensively, models enable automatic image classification and
capturing the relationships between nodes and the structural abnormal detection and assist physicians in diagnostic deci-
information of the graph. By leveraging the node relation- sion-making by integrating multimodal data. Thus, the rate
ships and information propagation within the graph struc- of misdiagnosis and missed diagnoses is significantly reduced,
ture, GCNs enhance the ability to process complex structured resulting in improved screening efficiency.
data. GCNs are increasingly used to interpret high-­resolution Finally, multiple performance metrics were used to evaluate
colposcopic images. Annotating and interpreting medical the AI model to ensure that the effectiveness was assessed in a
images requires well-trained cytologists, pathologists, and comprehensive manner. Model performance was assessed with
specialists with at least 5–10 years of experience, making the respect to sensitivity and specificity, accuracy, precision, recall,
process both time-consuming and resource-intensive. At pres- F1-score, and the area under the receiver operating character-
ent, deep learning algorithm exploration is primarily aimed at istic curve (AUC-ROC) as primary metrics. Intersection over
alleviating this issue. A semi- or weakly-supervised learning union (IoU), dice coefficient, and mean average precision
method, for example, can analyze and learn features from par- (mAP) are typically used to evaluate segmentation accuracy.
tially or minimally annotated images, applying pseudo-­labels Average precision (AP) was calculated for each class and the
to ­unannotated images for classification and object detection mAP was computed to provide an overall performance meas-
tasks. With self- and un-supervised learning methods, which ure across all classes. Additionally, evaluation metrics can be
do not require manually annotated category labels, feature used to evaluate the performance of classification algorithms.
learning can be achieved through a vast collection of unan- Metrics, such as diagnostic accuracy, which encompasses the
notated image samples. For example, generative adversarial overall accuracy of the AI diagnostic outputs, and positive pre-
networks (GANs) generate high-quality synthetic images, dictive values (PPVs) and negative predictive values (NPVs),
enhancing the diversity of datasets. GANs consist of a genera- indicating the likelihood of true positives and true negatives,
tor and discriminator that work through an adversarial training respectively, were also used to evaluate the AI model as a diag-
process to produce high-quality images. The generator creates nostic tool. These metrics are essential for assessing the prac-
realistic images to deceive the discriminator, while the discrim- tical applicability of the AI model in clinical settings and the
inator distinguishes between real and generated images. The potential impact on screening outcomes. The performance of
generator progressively improves the quality of the generated AI technology in the early screening and detection of cervical
images in this adversarial process, making it increasingly dif- cancer has been validated by several studies and has demon-
ficult for a discriminator to differentiate the generated images strated good diagnostic accuracy14-16.
from real images. The training method does not require labeled AI model workflows require rigorous data quality and
data and ensures the robustness of the model as well. Cervical model selection to avoid biases and noise that can cause
868 Wu et al. Artificial intelligence strengthens cervical cancer screening

under- or over-fitting, which affects the generalization abil- BestCyte also incorporates cell annotation and WSI review
ity of the model. A model is then selected and fine-tuned to through a remote operation platform for peer review by cyto-
achieve an optimal model based on performance. The high pathologists. These automated screening systems, however,
sensitivity and specificity are typically expected by clinical are not actually AI-assisted screening technologies, but rather
decision applications, but clinically acceptable results may forms of computer-aided imaging techniques.
vary depending on the application. Thus, internal and external Automated smear analysis involves the following pro-
validation is crucial to ensuring that the model is stable and cedures: digital image slide acquisition; identifying ROIs;
generalizable. In general, AI is an evolving process that will segmenting to isolate relevant features of cells; and classify-
require regular updates on datasets used to train the model as ing images into pre-neoplastic categories for cytopathologist
well as optimization for clinical use. review26. AI technology is primarily utilized in the segmen-
tation and classification phases, which helps reduce the daily
workload of cytopathologists and improves the efficiency of
AI-guided methods in enhancing screening. Figure 1 illustrates the workflow of deep learning
cervical cytology networks used in cervical cytology diagnosis. Segmentation
refers to the isolation of multiple regions of cells to extract pre-
An AI-assisted system analyzes cytologic images, develop cise information about the ROI for the detection of abnormal
mathematical models based on deep learning or other AI tech- cells. Generally, cytopathology requires a closer examination
niques, and screen digital smear images to identify normal and of characteristics of the cell nucleus, therefore the nucleus and
abnormal cells and facilitate cervical cancer screening. The cytoplasm must be precisely segmented. Sompawong et al.27
initial attempt to automate cervical smear screening system developed a Mask-RCNN architecture that used a classifi-
began in 1992 with approval of the PAPNET Testing System cation branch to distinguish between normal and abnormal
for rescreening of conventional cervical smears that were man- features based on the nuclear locations and a segmentation
ually screened17. The liquid-based cytology technique was an branch to pinpoint the nuclei locations on a Pap smear slide.
important innovation to the traditional Pap smear for improv- U-Net is a convolutional network widely used for medical
ing the quality of cervical cytology. The cervical samples are segmentation tasks. U-Net was introduced by Ronneberger
collected in liquid media (ThinPrep18 or SurePath19). After et al.28 as a model that can learn from a few annotated images.
preparation and staining of smears, the smears are scanned Several studies in recent years have examined the performance
with a microscope slide scanner to produce digital images. of U-net on the segmentation of cervical cells with a dice
The US Food and Drug Administration (FDA) approved two rate > 90% on internal datasets29-31. Zhang et al.30 validated
computer-aided imaging systems for automated detection of the proposed GC-UNet in an actual cervical cancer diagno-
abnormal cells in 2010 (BD FocalPoint GS Imaging System and sis setting, achieving a remarkable precision rate of 99.5%.
the ThinPrep Imaging System20,21). Although suspicious cells This finding indicates that U-net might be a highly effective
are detected on a slide by computer-aided systems, the entire method for segmenting cervical nuclei, which will serve as an
slide must be manually screened and interpreted by a cyto- important tool for diagnosing and screening cervical cancer.
pathologist. Although this system has improved with respect Additionally, the method shows significant potential for prac-
to sensitivity and efficiency21-23, extensive manual screening tical application, with its rapid processing time of just 0.85 sec-
is required and the final diagnosis is fully dependent upon onds per image. As compared to the U-net, Wang et al.32 pro-
the final cytopathologist manual screening process. BestCyte, posed a multi-layer deep learning framework for improving
a whole slide image (WSI)-based scanning technology, was the accuracy of detection of CIN2+ cells. The method employs
introduced in 201424. A powerful method is applied to catego- a coarse-to-fine strategy for quick identification of target ROI
rize and systematically display images of clinically significant tissue location through semantic segmentation, followed by
cells in galleries based on the cytomorphologic characteristics precise single HSIL cell detection on ­specific ROIs. The mul-
within fields of view (FOVs). BestCyte supports the anno- ti-layer deep learning framework is also 20 times faster than
tation of images at the cell level (40× magnification) could U-net in processing one piece of WSI.
potentially standardize objectivity among c­ytologists, lead- In general, the classification of cervical cells is intended to
ing to fewer discrepancies in final diagnoses25. In ­addition, improve detection rates of cervical intraepithelial neoplasia.
Cancer Biol Med Vol 21, No 10 October 2024 869

A WSI level

B Patch level Cell Cell

C segmentation D classification
LSIL E WSI diagnosis

HSIL
RoI
detection

ASC-H
Feature map
ASCUS
Feature extraction
...

LSIL HSIL ASCUS ASC-H

Image preprocessing Image analysis process Final diagnosis process

Figure 1 Schematic representation of AI-assisted cervical cytology image analysis. (A) Whole slide image (WSI) level: Digitalization of cervical
liquid-based preparation samples; (B) Patch level: WSIs are divided into smaller patches to create feature maps, focusing on significant cellular
structures and detects regions of interest (ROIs); (C) Cell segmentation: Segmentation isolates nuclei from each cell, emphasizing morphologic
features; (D) Cell classification: The extracted features classify cells into categories, such as LSIL, HSIL, ASC-H, and ASCUS; (E) WSI diagnosis:
The classification results are aggregated to provide an overall diagnosis at the WSI level.

Current AI-assisted slide recognition and cytologic classi- analysis platform based on deep learning and validated the
fication tasks rely on digital images from whole slide imag- platform in a large sample population. Compared with manual
ing, mainly including scanned conventional Pap smears and screening, the performance of AI classifier has similar sensi-
scanned LBC slides after staining. Recent studies have demon- tivity and higher specificity42. The simultaneous presence of
strated that deep learning methods, such as CNNs and feature p16 and Ki-67 in the same cell provides a valuable triage strat-
attention networks, can achieve an accuracy > 90% in both egy, thus this AI classifier reduces the workload of cytologists
binary and multiple classification tasks for conventional smear and saves unnecessary resources. Considering the variety of
images33-37. It is worth mentioning that Wang et al.33 contrib- types of cytologic images, Cheng et al.43 enhanced abnormal
uted a transfer-learned CNN, which is used for classification by cell detection in cervical smears by integrating CNN with a
a limited number of Pap smear images with coarse image-level recurrent neural network (RNN), providing an adaptable
labels but has achieved remarkable performance. However, algorithm that could be used with different slide preparations,
the liquid-based cytologic smear preparation method has the staining, and imaging methods. The most recent research on
advantage of a smaller scan area and minimizing obscurations AI systems for cytology is shown in Table 1.
and cell overlap, which has led to the development of efficient Aside from AI research on cervical cytology images,
AI cytology applications14,38-40. Bao et al.41 developed a super- AI-based digital microscopes have also been developed, pro-
vised deep learning algorithm based on 188,542 digital LBC viding new opportunities to address the challenges of cervical
images and evaluated the capability of detecting CIN2+ and cancer screening in LMICs. Tang et al.48 utilized the augmented
CIN3+ lesions. The accuracy of the algorithm was comparable reality (AR) technique with an AI microscope to provide
to that of experienced cytologists. The exploratory develop- real-time assistance for cervical cytology diagnosis. The AR
ment of deep learning classifiers also covers specially stained technique significantly improved the sensitivity of LSIL and
cytology slides, such as p16/Ki-67 dual-stained (DS) and HSIL, and also enhanced the consistency of various atypical
H&E-stained slides. Wentzensen et al.42 applied two neural squamous cells48. In addition, a digital diagnostic system for
networks (CNN4 and Inception-v3) to develop a new image Papanicolaou smears was developed by Holmstrom et al.49
870 Wu et al. Artificial intelligence strengthens cervical cancer screening

Table 1 AI system advances in cytology for cervical cancer

Author and Ref. Aim of study Number of subjects AI system performance Key study outcomes
Du H44 Feasibility and efficiency of 5,000 high-confidence Accuracy of NILM, HSIL, ASU, The interpretation time for
cytology slide interpretation slides and LSIL prediction on single each slide was reduced from
(private dataset) cells is 81.4%, 90%, 42.54%, 3 min to 30 seconds
and 68.23%, respectively

Bai X45 Identification and 32,451 cases Sensitivity of CIN2+ smear The average reading time of
interpretation on CINII and (private dataset) pap is 99.3% and specificity pathologists with AI system
above cervical smear pap 9.87% by AI alone was 22.23 seconds per case
compared to a manual
reading time of 180 seconds

Xue P15 The performance of an 489 cases The sensitivity of AI system at Compared to HPV16/18
AI-enabled liquid-based (private dataset) detecting CIN2+ is 86.49%, typing the AI system
cytology as a screening and the specificity is 51.33% sensitivity is substantially
triage approach higher and specificity
is lower. The AI system
reduced referrals
to colposcopy by
approximately 10%

Xue P46 The efficiency of abnormal 8,000 digitalized whole The sensitivity of AI alone is Reduced the cytology
cervical squamous cell slide images 89.4% and the specificity is workload by more than
detection in cervical cancer (private dataset) 66.4% one-third. The AI system
screening had superior sensitivity and
specificity compared to
junior cytologists

Bao H14 AI-assisted cytology system 703,103 cases The sensitivity of the AI The agreement rate
at different CIN levels of (private dataset) system on CIN1+, CIN2+, and between AI and manual
detection CIN3+ is 88.9%, 90.1%, and reading was 94.7%, which
90.9%, respectively; specificity was a 5.8% increase in
on CIN1+, CIN2+, and CIN3+ sensitivity compared to
is > 90% manual reading

Zhu X47 Classified cervical liquid- 34,403 smear samples The sensitivity of Achieving a speed < 180s/
based thin-layer cell smears (private dataset) intraepithelial lesions is 92% slide with high sensitivity;
on 5 classes and the specificity is 84.39% the sensitivity of senior
cytologists detection is
lower than the AI system

Wentzensen N42 Detection on dual-stain+ Based on 3 epidemiologic The sensitivity of CIN3+ cells The AI system was
cells and performance of AI studies, > 4,000 cases on AI DS-cytology (single developed using P16/
cytology in cervical cancer (private dataset) cell) is 91.8% Ki-67 dual-staining
screening slides; AI-based cytology
interpretation is more
sensitive than manual; AI
results reduced colposcopy
referrals by one-third

HSIL, high-grade squamous intraepithelial lesion; LSIL, low-grade squamous intraepithelial lesion; NILM, negative for intraepithelial lesion or
malignancy; ASU, atypical squamous cells of undetermined significance; CIN, cervical intraepithelial neoplasia; AI, artificial intelligence.

using a portable whole-slide microscope scanner and a deep in high sensitivity, especially for detecting high-grade atypia
convolutional network trained on commercially available slides, which might help reduce the workload for micros-
image-analysis platforms for the detection of cervical carci- copists and cytopathologists in low-resource settings. The
noma cells4. The Holmstrom et al. study49 found that using the findings of all the research studies validate the potential and
system for identifying squamous cell atypia is viable, resulting benefits of using AI to aid in cytologic diagnosis, as well as
Cancer Biol Med Vol 21, No 10 October 2024 871

the potential to assist in screening efforts in areas with lim- lesions from annotated colposcopic images with assessment
ited resources. Moreover, Tang et al.48 discussed the poten- of suspicious lesion areas using image recognition algo-
tial utilization of AI technology, such as AI microscopes, to rithms. Due to the subjectivity of colposcopic diagnosis,
enhance the professional training of newly trained cytopathol- AI technology is critical in helping primary care colposco-
ogists in low-resource settings. Hologic, Inc. announced the pists in low-resource healthcare areas correctly differentiate
launch of the first FDA-approved digital cytology system in between normal and abnormal cervical findings, grading,
February 2024 (Genius™ Digital Diagnostics System equipped and categorizing cervical lesions efficiently. Few studies
with Genius™ Cervical AI algorithm and volumetric imaging have evaluated the effectiveness of AI-based diagnosis with
technology). This diagnostics system consists of image acqui- smartphone-obtained colposcope images, and these studies
sition, analysis of images by the AI algorithm, image storage, have shown promising results that are systematically supe-
and remote peer review50. The Genius™ Digital Diagnostics rior to t medical experts52-55. AI technology is increasingly
System demonstrates that AI technology has a bright future being used to assist experienced colposcopists in enhancing
in cervical cytology and is projected to have a significant their diagnostic performance, classifying lesions more effec-
impact on cervical cancer screening during the coming years. tively, identifying the transformation zone (TZ), and guiding
AI-assisted liquid-based cytology testing may facilitate the colposcopists in determining biopsy sites (Figure 2). The
rapid expansion of cervical cancer screening, while also being most recent AI-colposcope research has concentrated on
more cost-effective51. the development of deep learning-based classifiers for cer-
vical neoplasia on magnified cervix images obtained with
special equipment, which increases the consistency with
AI applications in colposcopic histopathologic findings56-63. Considering the heterogeneity
diagnosis and assistance in biopsies of colposcopic imaging equipment and the prevalent lack of
standardized annotations of colposcopic images, the appli-
The AI-assisted colposcopy diagnostic system combines cation of semi-supervised learning algorithms for inferring
high-definition colposcopic imaging to identify cervical cervical dysplasia categorization from limited high-quality

Figure 2 Illustrative example of an AI-assisted colposcopy-guided biopsies for a case diagnosed as HSIL/CIN2.
872 Wu et al. Artificial intelligence strengthens cervical cancer screening

colposcopy images represents a current research trajectory developing diagnostic abilities to the level of experienced prac-
within the field of AI-assisted colposcopy64,65. Based on the titioners and guide novice colposcopists in performing effi-
ASCCP colposcopy standards, one of the most critical fac- cient biopsy procedures. Given these findings, an AI-assisted
tors for grading colposcopic findings is the type of cervical colposcopy diagnostic system will have a valuable support role
TZ and whether it is fully or partially visible66. Referencing in cervical cancer screening in areas with limited resources. AI
the Colposcopy Terminology published by the International colposcopy systems, like CAIADS and Cerviray AI, are also
Federation for Cervical Pathology and Colposcopy (IFCPC), equipped with remote access functionality. Thus, the further
a cervical TZ is typically defined as a region where squamous development of cloud-based AI colposcopy platforms might
metaplasia has developed and is known to be a predispos- narrow the gap in colposcopic examinations between LMICs.
ing site for cervical cancer development47. Thus, few studies In addition, mobile-based and other portable hardware col-
have been conducted on the implementation of deep learn- poscopy devices that incorporate AI technology in a more
ing algorithms to improve the segmentation of the acetow- accessible and user-friendly format, such as MobileODT77
hite lesion and determine its TZs67,68. The results of these and Cervicare AI78, have also shown promising performance
studies reveal that precise segmentation of the TZ can effec- in validation studies and have been successfully commercial-
tively enhance the discriminative representation capacity of ized. However, there is currently no FDA approved AI-based
the deep learning-based CIN classifier60,63. In clinical colpos- colposcopy tool for cervical cancer detection. Currently, AI in
copy practice, a critical objective of cervical cancer screen- colposcopy aims to address the shortage of experienced col-
ing is differentiating CIN grades. When lesions are detected, poscopists, reduce misdiagnosis rates, and enhance the diag-
biopsies of 2–4 sites are obtained to ascertain the most severe nostic efficiency of traditional colposcopy. In fact, AI holds
lesion. For lesions diagnosed as CIN 2/3, treatments, such as significant potential for providing effective education in col-
conization or LEEP, might be required. To improve the accu- poscopy. The IARC offers training resources for colposcopy
racy and appropriateness of biopsy sites, some AI-assisted in several languages, providing comprehensive and accessible
colposcopes provide guidance regarding cervical biopsies educational materials to a global audience79. A recent study by
and predict the location of the biopsy site57,69,70. The most Chen et al.80 developed an online digital education tool with
current advanced AI models for application on colposcopy numerous real-world colposcopy images for colposcopy train-
are summarized in Table 2. ing, which provided short-term improvements in colposcopist
The development of AI models for the classification of cer- competency and confidence. In addition to the acquisition
vical lesions has resulted in impressive results, even achieving of standard terminology as well as a greater understanding
diagnostic capabilities comparable to colposcopists in some of colposcopy, digital training platforms facilitate interactive
studies69,74-76. However, the independent interpretability of educational exchanges with colposcopists. This platform ena-
AI models still lacks clinical credibility. Moreover, very few of bles learners to customize their enrichment of knowledge by
these developed AI models have been validated for applicabil- addressing specific educational needs (Figure 3). Based on this
ity to real clinical use. Kim et al.72 evaluated the feasibility of inspiration, AI interpretability on colposcopic images might
the Cerviray AI system® on 234 patients and reporterrd supe- also be a potential training benefit for novice colposcopists.
rior sensitivity and similar specificity over two colposcopists
for detecting high-grade lesions. However, the sensitivity sig-
nificantly improved when an AI system worked in conjunction
AI-assisted Cervical Cancer
with at least one colposcopist. Wu et al.71 conducted a retro- Screening Challenges and
spective hospital-based study evaluating the colposcopic AI Suggestions
diagnostic system, CAIADS. CAIADS guided fewer biopsy sites
and had the greatest biopsy sensitivity for high-grade lesions AI will have significant implications for cervical cancer screen-
compared to subspecialists71. With the assistance of CAIADS, ing, especially with AI-based methods for cytology screening
the sensitivity achieved by junior colposcopists on CIN grades that have achieved considerable technical proficiency. This
and biopsy was significantly improved. These studies demon- has been attributed to the development of AI algorithms in
strated the clinical applicability of the AI-assisted colposcopy cervical cytology over the past few years. The precision of
diagnostic system, which can assist novice colposcopists in segmentation models29,31,32,81,82 and accuracy of classification
Cancer Biol Med Vol 21, No 10 October 2024 873

Table 2 Application of AI model in colposcopy to detect cervical cancer

Author and Ref. Aim of study Study type Number of subjects Key study outcomes
Yuan C60 Detection on LSIL+ Model development 22,330 cases for AI model The AI model was able to segment and
colposcopy and validation study training and evaluation; classify LSIL and HSIL cervical lesions. The
5384 cases for validation accuracy of the AI model on LSIL is 84.1%
(private dataset) and the sensitivity of the AI model on HSIL is
88.47%. In the validation study, 84.67% HSIL
cases were detected, which was better than
the colposcopist

Yan L61 Detection on LSIL/HSIL Model development 7,530 patients, Significant 95% accuracy on normal/LSIL
colposcopy study 15,276 images classification, and 90% accuracy on HSIL-and
(private dataset) HSIL+; a stronger diagnostic performance
than the junior colposcopist in 300 samples
from the test set

Xue P69 Detection on LSIL/HSIL Model development Total 19,435 patients and Accuracy of AI model on LSIL/HSIL
colposcopy and validation study 101,267 images classification is 80.7% compared to the
(private dataset) colposcopist interpretation on the validation
dataset. The AI model showed slightly higher
sensitivity: 65.8% vs. 60.4%

Wu A71 Performance of Hospital-based 1,146 patients The average sensitivity of CAIADS on

CAIADS69 on CIN2+/ retrospective study (private dataset) CIN2+/CIN3+ is 80%, which was not lower
CIN3+ detection (AI model external than a senior colposcopist. The sensitivity
study) of the junior colposcopist with CAIADS is
significantly improved. Number of biopsies
recommended by CAIADS per case was less
than the colposcopist

Kim S72 Evaluation on the Observation 234 patients The final diagnostic accuracy of Physician
feasibility of interpreting study (AI system (private dataset) 1 with AI-assist on colposcopy images
colposcopy images with application study) increased from 76% to 80%, and the accuracy
the AI-assist of Physician 2 increased from 71% to 77%

Fu L13 Improved colposcopy Model developmet 2,160 cases The diagnostic performance with AUC
DL-base model with and validation study (private dataset) was improved to 0.921 as a multimodal
HPV test result and integrated model from AUC 0.84 as a
cytology test result colposcopy-based DL model

Mukku J73 CIN detection on Model development 900 images from IARC Impressive 89.32% sensitivity and 91.6%
colposcopy images with study image bank11 specificity on diagnosing CIN with fusion
clinical outcomes by strategy on various clinical findings
multimodal strategy (including age, HPV test, biopsy result, and
transformation zone)

HSIL, high-grade squamous intraepithelial lesion; LSIL, low-grade squamous intraepithelial lesion; CAIADS, colposcopic artificial intelligence
auxiliary diagnostic system; CIN, cervical intraepithelial neoplasia; AI, artificial intelligence; DL, deep learning.

­ odels33-35,37 have demonstrated impressive evaluation results,

m datasets to train different segmentation and classification tasks.
with both exceeding 90% in precision and accuracy. From the Additionally, the development of AI-based colposcopy diag-
perspective of learning strategy, there has been a shift from nostic models still requires further enhancement to ensure pre-
ensemble learning83,84 to transfer learning33,35. Advances in AI cise cervical lesion classification, including increasing the spec-
algorithms for cytology have been enhanced by several public ificity of early cervical lesions and accurately subdividing more
datasets, including the Herlev9, ISBI Challenge6,7, Sipakmed8, subtypes of cervical lesions (e.g., adenocarcinoma in situ and
and Cx22 databases5. Various types of Pap smear, single-cell, cervical adenocarcinoma). In general, AI algorithm models are
and overlapping cervical cell images are provided in these accurate to approximately 85% in the classification of cervical
874 Wu et al. Artificial intelligence strengthens cervical cancer screening

A Learning Progress Chapter 6 A systematic approach to colposcopic

examination
Consolidation Practice
Basic Knowledge of Colposcopy
From «COLPOSCOPY AND TREATMENT OF CERVICAL
30% PRECANCER »
10 cases/estimated 60 mins
Page 1/Total 7 Continue learning Continue
Beginner Goal

B Recommended Learning Self-assessment

Most viewed
Ectropion/eversion of columnar epithelium Come and test yourself
Ectropion or eversion refers to the displacement of columnar
epithelium from within the cervical canal to the exterior of
the ectocervix. Provide customized learning program
Go to test >

Recommended for “Beginner Goal” Assessment Records

How to determine the TZ types?

05:49

C Stage Exercise Knowledge Plaza

Beginner Intermediate Advanced Official

Guidelines Terminology
Practice Practice Practice Textbook

Figure 3 Sample interface of an AI-guided colposcopy training platform. The platform offers a structured learning, using AI to tailor content
and enhance colposcopy training. (A) Displays the user learning progress and access to textbook chapters and consolidation practice exer-
cises. (B) Provides personalized learning materials based on self-assessment results through a recommendation system. (C) Provides exercises
for beginner, intermediate, and advanced levels. Includes access to guidelines, official textbooks, and terminology resources.

lesions on colposcopy images. It is interesting to note that light- (LLMs), the revolution in artificial intelligence has officially
weight neural networks, such as MobileNet, EfficientNet, and begun. Recently, Meta’s FAIR lab Segment Anything Model
SqueezeNet, have increasingly been adopted as backbone net- (SAM), a giant AI image segmentation model aiming to revolu-
works in both cytology and colposcopy62,67,85. Given this trend, tionize machine learning for versatile and accurate image seg-
equally efficient models can likely be deployed on mobile mentation, was released86. This state-of-the-art model will not
devices or portable computers, making the models more suit- be restricted by image types and domains. Wang et al.87 devel-
able for environments that are limited in resources. However, oped a foundational model based on whole-slide pathologic
there are very few publicly available colposcopy image data- images, demonstrating that the latest generative AI technolo-
sets, with the IARC Cervical Cancer Image Bank11 being one gies are already being applied in the medical field to address the
of the few that meet expert qualification standards for collec- modeling challenges of such large-scale medical images. This
tion procedures and image quality. Most colposcopy images general model can also be applied to other types of image data,
used in AI algorithm development are not publicly available. such as CT, MRI, or X-rays. Therefore, for the development of
Thus, establishing a standardized and regulated data platform advanced AI algorithm models, processing colposcopy images
for systematic management and quality control is essential for from different devices of varying qualities by one foundation
further advances. With the emergence of large language models AI model will be a future direction.
Cancer Biol Med Vol 21, No 10 October 2024 875

As discussed in this review, validation studies that evaluate every 5 y, the cost-effectiveness could be comparable to HPV
AI models may show heterogeneity due to factors, such as the testing51. AI systems can significantly reduce diagnostic time
diversity of populations, variations in slide preparation, and and cost. However, pathologists and colposcopists using these
various evaluation metrics. Therefore, it is difficult to provide systems to assist with diagnostic purposes must navigate the
a direct comparison of AI algorithms. Most evaluation results allocation of responsibility for clinical outcomes, as well as
indicate that AI models used in clinical validation for cytology the privacy and security of patient data, which require further
have difficulty achieving a 0.9 sensitivity14,15,46, although the definition. Hence, AI applications can only assist physicians
speed of slide reading has significantly improved45,47. When in diagnosing rather than replace physicians in clinical deci-
compared to cytologists, AI-based systems for CIN detection sion-making. Healthcare professionals should understand the
are generally in line with the proficiency of junior cytologists, advantages and limitations of AI when using AI-based prod-
with some exceeding senior cytologists in performance14,47. ucts. Furthermore, AI must be used in a manner that respects
Additionally, the specificity of different AI models varies con- the autonomy of patients. To ensure ethical compliance with
siderably, most likely due to the varying case sample distribu- AI, it is imperative that clear guidelines and regulations be
tions included in the studies. Despite the relatively advanced established as well. Digital health innovation is a focus of
application of AI image analysis in cytology4, increasing the US FDA Digital Health Innovation Action Plan88, which
clinical validation studies and standardizing the method of is intended to regulate and monitor digital health devices.
reporting the results of these studies should be prioritized. Similarly, the European Union Medical Device Regulation
Furthermore, large-scale studies are lacking for the clinical emphasizes stringent requirements for the clinical evalua-
validation of AI image analysis in colposcopy. Currently, only tion and post-market surveillance of AI-powered medical
Yuan et al.60 and Xue et al.69 have included sufficient num- devices89. The National Medical Products Administration
bers of cases to evaluate the performance of AI systems in in China is strengthening its regulatory processes to accom-
clinical settings. However, despite the improved sensitivity of modate rapid advances in AI technology. The International
AI systems compared to junior colposcopists in these valida- Telecommunication Union is working on international stand-
tion studies61,71,72, some differences remain. An advantage of ards to facilitate global harmonization90. Global efforts are
AI-assisted colposcopy is that it facilitates guided biopsies and being made towards establishing robust governance for AI in
enables women to detect lesions more effectively with fewer healthcare, ensuring patient safety, and promoting innovation.
biopsy samples, thereby causing less cervical damage. Yuan This technology must be approached with caution and strict
et al.60 showed that the number of biopsies performed using regulations in place to protect patient privacy, informed con-
the AI system was slightly higher than the number performed sent, and ethical considerations to prevent misuse.
by the colposcopist in each case. However, Wu et al.71 reported
that the AI system performed fewer biopsies than colposco- Conclusions
pists. The significant variability in colposcopy image capture
and collection results in AI system performance differences, In this paper we have described the current development,
emphasizing the importance of standardized protocols in application, challenges, and future directions of AI in cervical
colposcopy image collection. Additionally, there will be an cancer screening. Cervical cancer screening methods with AI
emphasis on assessing the extent to which AI models can be technology have the potential to significantly transform the
applied to large real-world populations based on prospective prevention and control of cervical cancer. Further applications
clinical studies to improve the effectiveness of cervical cancer of AI to cervical cancer screening could deliver high-quality
screening, further incorporating screening strategies. clinical performance, provide diagnostic rationale of expla-
Finally, despite the potential benefits and cost-effectiveness nation and interpretability on standardized platforms, and
of AI in cervical cancer screening, clinicians are concerned archive extensive real-world cervical images for education
that the lack of clear interpretability of diagnostic decisions purposes. This might enable the gap between tertiary and pri-
raises substantial concerns regarding safety, resilience, and mary care hospitals to be narrowed, in turn maximizing health
ethical considerations. If AI technology is integrated into care for a broader segment of the population. A further appli-
cervical cancer screening strategies, especially by using cation of AI will be in the prevention and control of cervical
AI-assisted cytology to replace traditional cytology screening cancer, reducing the workload of medical personnel while
876 Wu et al. Artificial intelligence strengthens cervical cancer screening

increasing diagnostic accuracy and efficiency. It is hoped that 6. Lu Z, Carneiro G, Bradley AP. An improved joint optimization of
current research on AI is expected to translate into clinical multiple level set functions for the segmentation of overlapping
cervical cells. IEEE Trans Image Process. 2015; 24: 1261-72.
practice, which will expedite the global goal of eliminating
7. Lu Z, Carneiro G, Bradley AP, Ushizima D, Nosrati MS, Bianchi
cervical cancer. AG, et al. Evaluation of three algorithms for the segmentation of
overlapping cervical cells. IEEE J Biomedical Health Inform. 2016;
Acknowledgements 21: 441-50.
8. Plissiti ME, Dimitrakopoulos P, Sfikas G, Nikou C, Krikoni O,
We thank Dr. Peng Xue and Ms. Mingyang Chen at Chinese Charchanti A. SIPaKMeD: a new dataset for feature and image
based classification of normal and pathological cervical cells in
Academy of Medical Sciences and Peking Union Medical
pap smear images. In: 2018 25th IEEE International Conference on
College for comments of the original manuscript. Image Processing (ICIP). 2018. p. 3144-8.
9. Jantzen J, Norup J, Dounias G, Bjerregaard B. Pap-smear
Grant support benchmark data for pattern classification. In: Nature Inspired
Smart Information Systems (NiSIS 2005). 2005. p. 1-9.
10. Payette J, Rachleff J, de Graaf C. Intel and MobileODT cervical
This study was supported by grants from CAMS Innovation
cancer screening Kaggle competition: cervix type classification
Fund for Medical Sciences (Grant No. CAMS 2021-I2M-1-
using deep learning and image classification. Stanford University;
004) and from the Bill & Melinda Gates Foundation (Grant 2017.
No. INV-031449). 11. IARC – International Agency for Research on Cancer. IARC visual
inspection with acetic acid (via) image bank. https://screening.iarc.
Conflicts of interest statement fr/cervicalimagebank.php.
12. Li J, Hu P, Gao H, Shen N, Hua K. Classification of cervical lesions
based on multimodal features fusion. Comput Biol Med. 2024; 177:
No potential conflicts of interest are disclosed.
108589.
13. Fu L, Xia W, Shi W, Cao G-X, Ruan Y-T, Zhao X-Y, et al. Deep
Author contributions learning based cervical screening by the cross-modal integration of
colposcopy, cytology, and HPV test. Int J Med Inform. 2022; 159:
Conceived and designed the analysis: Partha Basu, Youlin 104675.
14. Bao H, Sun X, Zhang Y, Pang B, Li H, Zhou L, et al. The artificial
Qiao, Fanghui Zhao.
intelligence-assisted cytology diagnostic system in large-scale
Collected the data: Tong Wu.
cervical cancer screening: a population-based cohort study of 0.7
Performed the analysis: Tong Wu, Eric Lucas. million women. Cancer Med. 2020; 9: 6896-906.
Wrote the paper: Tong Wu, Partha Basu. 15. Xue P, Xu HM, Tang HP, Wu WQ, Seery S, Han X, et al. Assessing
artificial intelligence enabled liquid-based cytology for triaging
HPV-positive women: a population-based cross-sectional study.
References Acta Obstet Gynecol Scand. 2023; 102: 1026-33.
16. Hu L, Bell D, Antani S, Xue Z, Yu K, Horning MP, et al. An
1. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, observational study of deep learning and automated evaluation of
et al. Global cancer statistics 2022: GLOBOCAN estimates of cervical images for cancer screening. J Natl Cancer Inst. 2019; 111:
incidence and mortality worldwide for 36 cancers in 185 countries. 923-32.
CA Cancer J Clin. 2024; 74: 229-63. 17. Mango LJ. Computer-assisted cervical cancer screening using
2. Zhao F, Qiao Y. Cervical cancer prevention in China: a key to neural networks. Cancer Lett. 1994; 77: 155-62.
cancer control. Lancet. 2019; 393: 969-70. 18. Lee KR, Ashfaq R, Birdsong GG, Corkill ME, Mcintosh KM,
3. Nazari Z, Torabizadeh G, Khalilian A, Ghadami S, Karimi- Inhorn SL. Comparison of conventional Papanicolaou smears and a
Zarchi M, Allahqoli L, et al. Is cryotherapy effective in all women fluid-based, thin-layer system for cervical cancer screening. Obstet
with low-grade cervical intraepithelial neoplasia? Eur Rev Med Gynecol. 1997; 90: 278-284.
Pharmacol Sci. 2021; 25: 4211-8. 19. Fremont-Smith M, Marino J, Griffin B, Spencer L, Bolick D.
4. Hou X, Shen G, Zhou L, Li Y, Wang T, Ma X. Artificial intelligence Comparison of the SurePath™ liquid-based Papanicolaou smear
in cervical cancer screening and diagnosis. Front Oncol. 2022; 12: with the conventional Papanicolaou smear in a multisite direct-to-
851367. vial study. Cancer. 2004; 102: 269-79.
5. Liu G, Ding Q, Luo H, Sha M, Li X, Ju M. Cx22: a new publicly 20. Chivukula M, Saad RS, Elishaev E, White S, Mauser N, Dabbs DJ.
available dataset for deep learning-based segmentation of cervical Introduction of the Thin Prep Imaging System™ (TIS): experience
cytology images. Comput Biol Med. 2022; 150: 106194. in a high volume academic practice. Cytojournal. 2007; 4: 6.
Cancer Biol Med Vol 21, No 10 October 2024 877

21. Wilbur DC, Black-Schaffer WS, Luff RD, Abraham KP, Kemper C, 35. Cao L, Yang J, Rong Z, Li L, Xia B, You C, et al. A novel attention-
Molina JT, et al. The Becton Dickinson FocalPoint GS Imaging guided convolutional network for the detection of abnormal cervical
System: clinical trials demonstrate significantly improved cells in cervical cancer screening. Med Image Anal. 2021; 73: 102197.
sensitivity for the detection of important cervical lesions. Am J Clin 36. Hussain E, Mahanta LB, Das CR, Talukdar RK. A comprehensive
Pathol. 2009; 132: 767-75. study on the multi-class cervical cancer diagnostic prediction on
22. Bolger N, Heffron C, Regan I, Sweeney M, Kinsella S, McKeown M, Pap smear images using a fusion-based decision from ensemble
et al. Implementation and evaluation of a new automated deep convolutional neural network. Tissue Cell. 2020; 65: 101347.
interactive image analysis system. Acta Cytol. 2006; 50: 483-91. 37. Bhatt AR, Ganatra A, Kotecha K. Cervical cancer detection in pap
23. Biscotti CV, Dawson AE, Dziura B, Galup L, Darragh T, Rahemtulla smear whole slide images using convNet with transfer learning and
A, et al. Assisted primary screening using the automated ThinPrep progressive resizing. PeerJ Comput Sci. 2021; 7: e348.
Imaging System. Am J Clin Pathol. 2005; 123: 281-7. 38. Ke J, Shen Y, Lu Y, Deng J, Wright JD, Zhang Y, et al. Quantitative
24. Delga A, Goffin F, Kridelka F, Marée R, Lambert C, Delvenne P. analysis of abnormalities in gynecologic cytopathology with deep
Evaluation of CellSolutions BestPrep® automated thin-layer liquid- learning. Lab Invest. 2021; 101: 513-24.
based cytology Papanicolaou slide preparation and BestCyte® cell 39. Lin H, Chen H, Wang X, Wang Q, Wang L, Heng P-A. Dual-path
sorter imaging system. Acta Cytol. 2014; 58: 469-77. network with synergistic grouping loss and evidence driven risk
25. Chantziantoniou N. BestCyte® primary screening of 500 ThinPrep stratification for whole slide cervical image analysis. Med Image
Pap Test thin-layers: 3 Cytologists’ Interobserver diagnostic Anal. 2021; 69: 101955.
concordance with predicate manual microscopy relative to Truth 40. Lasyk Ł, Barbasz J, Żuk P, Prusaczyk A, Włodarczyk T, Prokurat E,
Reference diagnoses defining NILM, ASCUS+, LSIL+, and ASCH+ et al. An evaluation of the construction of the device along with the
thresholds for specificity, sensitivity, and equivalency grading. software for digital archiving, sending the data, and supporting the
J Pathol Inform. 2023; 14: 100182. diagnosis of cervical cancer. Contemp Oncol. 2019; 23: 171-7.
26. Rezende MT, Bianchi AG, Carneiro CM. Cervical cancer: 41. Bao H, Bi H, Zhang X, Zhao Y, Dong Y, Luo X, et al. Artificial
automation of Pap test screening. Diagn Cytopathol. 2021; 49: intelligence-assisted cytology for detection of cervical
559-74. intraepithelial neoplasia or invasive cancer: a multicenter, clinical-
27. Sompawong N, Mopan J, Pooprasert P, Himakhun W, Suwannarurk based, observational study. Gynecol Oncol. 2020; 159: 171-8.
K, Ngamvirojcharoen J, et al. Automated pap smear cervical cancer 42. Wentzensen N, Lahrmann B, Clarke MA, Kinney W, Tokugawa D,
screening using deep learning. Annu Int Conf IEEE Eng Med Biol Poitras N, et al. Accuracy and efficiency of deep-learning-based
Soc. 2019; 2019: 7044-8. automation of dual stain cytology in cervical cancer screening.
28. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks J Natl Cancer Inst. 2021; 113: 72-9.
for biomedical image segmentation. In: Proceedings of the 18th 43. Cheng S, Liu S, Yu J, Rao G, Xiao Y, Han W, et al. Robust whole
International Conference on Medical Image Computing and slide image analysis for cervical cancer screening using deep
Computer-Assisted Intervention – MICCAI 2015, Munich, learning. Nat Commun. 2021; 12: 5639.
Germany, October 5-9, 2015. 2015. p. 234-41. 44. Du H, Dai W, Zhou Q, Li C, Li SC, Wang C, et al. AI-assisted
29. Rasheed A, Shirazi SH, Umar AI, Shahzad M, Yousaf W, Khan Z. system improves the work efficiency of cytologists via excluding
Cervical cell’s nucleus segmentation through an improved UNet cytology-negative slides and accelerating the slide interpretation.
architecture. PLoS One. 2023; 18: e0283568. Front Oncol. 2023; 13: 1290112.
30. Zhang E, Xie R, Bian Y, Wang J, Tao P, Zhang H, et al. Cervical 45. Bai X, Wei J, Starr D, Zhang X, Wu X, Guo Y, et al. Assessment of
cell nuclei segmentation based on GC-UNET. Heliyon. 2023; 9: efficacy and accuracy of cervical cytology screening with artificial
e17647. intelligence assistive system. Mod Pathol. 2024; 37: 100486.
31. Ji J, Zhang W, Dong Y, Lin R, Geng Y, Hong L. Automated cervical 46. Xue P, Xu H-M, Tang H-P, Weng H-Y, Wei H-M, Wang Z, et al.
cell segmentation using deep ensemble learning. BMC Med Improving the accuracy and efficiency of abnormal cervical
Imaging. 2023; 23: 137. squamous cell detection with cytologist-in-the-loop artificial
32. Wang C-W, Liou Y-A, Lin Y-J, Chang C-C, Chu P-H, Lee Y-C, et al. intelligence. Mod Pathol. 2023; 36: 100186.
Artificial intelligence-assisted fast screening cervical high grade 47. Zhu X, Li X, Ong K, Zhang W, Li W, Li L, et al. Hybrid AI-assistive
squamous intraepithelial lesion and squamous cell carcinoma diagnostic model permits rapid TBS classification of cervical liquid-
diagnosis and treatment planning. Sci Rep. 2021; 11: 16244. based thin-layer cell smears. Nature Commun. 2021; 12: 3541.
33. Wang P, Wang J, Li Y, Li L, Zhang H. Adaptive pruning of transfer 48. Tang HP, Cai D, Kong YQ, Ye H, Ma ZX, Lv HS, et al. Cervical
learned deep convolutional neural network for classification of cytology screening facilitated by an artificial intelligence microscope:
cervical Pap smear images. IEEE Access. 2020; 8: 50674-83. a preliminary study. Cancer Cytopathol. 2021; 129: 693-700.
34. Nambu Y, Mariya T, Shinkai S, Umemoto M, Asanuma H, Sato I, 49. Holmström O, Linder N, Kaingu H, Mbuuko N, Mbete J, Kinyua F,
et al. A screening assistance system for cervical cytology of et al. Point-of-care digital cytology with artificial intelligence for
squamous cell atypia based on a two-step combined CNN cervical cancer screening in a resource-limited setting. JAMA
algorithm with label smoothing. Cancer Med. 2022; 11: 520-9. Network Open. 2021; 4: e211740.
878 Wu et al. Artificial intelligence strengthens cervical cancer screening

50. Bounds OCKN. Sustainability report. https://www.hologic.com/ network. In: 2020 IEEE 17th International Symposium on
hologic-products/cytology/genius-digital-diagnostics-system. Biomedical Imaging (ISBI). 2020. p. 1720-4.
51. Shen M, Zou Z, Bao H, Fairley CK, Canfell K, Ong JJ, et al. Cost- 66. Khan MJ, Werner CL, Darragh TM, Guido RS, Mathews C,
effectiveness of artificial intelligence-assisted liquid-based cytology Moscicki A-B, et al. ASCCP colposcopy standards: role of
testing for cervical cancer screening in China. Lancet Reg Health colposcopy, benefits, potential harms, and terminology for
West Pac. 2023; 34: 100726. colposcopic practice. J Lower Genit Tract Dis. 2017; 21: 223-9.
52. Xue Z, Novetsky AP, Einstein MH, Marcus JZ, Befano B, Guo P, 67. Yue Z, Ding S, Li X, Yang S, Zhang Y. Automatic acetowhite lesion
et al. A demonstration of automated visual evaluation of cervical segmentation via specular reflection removal and deep attention
images taken with a smartphone camera. Int J Cancer. 2020; 147: network. IEEE J Biomed Health Inform. 2021; 25: 3529-40.
2416-23. 68. Liu J, Liang T, Peng Y, Peng G, Sun L, Li L, et al. Segmentation of
53. Vinals R, Vassilakos P, Rad MS, Undurraga M, Petignat P, Thiran acetowhite region in uterine cervical image based on deep learning.
J-P. Using dynamic features for automatic cervical precancer Technol Health Care. 2022; 30: 469-82.
detection. Diagnostics. 2021; 11: 716. 69. Xue P, Tang C, Li Q, Li Y, Shen Y, Zhao Y, et al. Development and
54. Zhang Y, Zall Y, Nissim R, Zimmermann R. Evaluation of a new validation of an artificial intelligence system for grading colposcopic
dataset for visual detection of cervical precancerous lesions. Expert impressions and guiding biopsies. BMC Med. 2020; 18: 1-10.
Systs Appl. 2022; 190: 116048. 70. Ueda A, Yamaguchi K, Kitamura S, Taki M, Yamanoi K, Murakami
55. Ito Y, Miyoshi A, Ueda Y, Tanaka Y, Nakae R, Morimoto A, et al. R, et al. Development of an artificial intelligence-based diagnostic
An artificial intelligence-assisted diagnostic system improves the system for the detection of abnormal colposcopic findings. JCO
accuracy of image diagnosis of uterine cervical lesions. Mol Clin Glob Oncol. 2023; 9: 82.
Oncol. 2022; 16: 1-6. 71. Wu A, Xue P, Abulizi G, Tuerxun D, Rezhake R, Qiao Y. Artificial
56. Peng G, Dong H, Liang T, Li L, Liu J. Diagnosis of cervical intelligence in colposcopic examination: a promising tool to assist
precancerous lesions based on multimodal feature changes. junior colposcopists. Front Med. 2023; 10: 1060451.
Comput Biol Med. 2021; 130: 104209. 72. Kim S, Lee H, Lee S, Song J-Y, Lee J-K, Lee N-W. Role of artificial
57. Cho B-J, Choi YJ, Lee M-J, Kim JH, Son G-H, Park S-H, et al. intelligence interpretation of colposcopic images in cervical cancer
Classification of cervical neoplasms on colposcopic photography screening. Healthcare. 2022; 10: 468.
using deep learning. Sci Rep. 2020; 10: 13652. 73. Mukku L, Thomas J. Multimodal early fusion strategy based
58. Li Y, Chen J, Xue P, Tang C, Chang J, Chu C, et al. Computer-aided on deep learning methods for cervical cancer identification. In:
cervical cancer diagnosis using time-lapsed colposcopic images. Congress on Intelligent Systems. 2023. p. 109-18.
IEEE Trans Med Imaging. 2020; 39: 3403-15. 74. Miyagi Y, Takehara K, Miyake T. Application of deep learning to
59. Luo Y-M, Zhang T, Li P, Liu P-Z, Sun P, Dong B, et al. MDFI: the classification of uterine cervical squamous epithelial lesion
multi-CNN decision feature integration for diagnosis of cervical from colposcopy images. Mol Clin Oncol. 2019; 11: 583-9.
precancerous lesions. IEEE Access. 2020; 8: 29616-26. 75. Li Y, Liu Z-H, Xue P, Chen J, Ma K, Qian T, et al. Grand: a large-
60. Yuan C, Yao Y, Cheng B, Cheng Y, Li Y, Li Y, et al. The application scale dataset and benchmark for cervical intraepithelial neoplasia
of deep learning based diagnostic system to cervical squamous grading with fine-grained lesion description. Med Image Anal.
intraepithelial lesions recognition in colposcopy images. Sci Rep. 2021; 70: 102006.
2020; 10: 11639. 76. Zimmer-Stelmach A, Zak J, Pawlosek A, Rosner-Tenerowicz A,
61. Yan L, Li S, Guo Y, Ren P, Song H, Yang J, et al. Multi-state Budny-Winska J, Pomorski M, et al. The application of artificial
colposcopy image fusion for cervical precancerous lesion diagnosis intelligence-assisted colposcopy in a tertiary care hospital within a
using BF-CNN. Biomed Signal Process Control. 2021; 68: 102700. cervical pathology diagnostic unit. Diagnostics. 2022; 12: 106.
62. Chen X, Pu X, Chen Z, Li L, Zhao KN, Liu H, et al. Application of 77. Mitchell EM, Doede AL, McLean Estrada M, Granera OB,
EfficientNet-B0 and GRU-based deep learning on classifying the Maldonado F, Dunn B, et al. Feasibility and acceptability of tele-
colposcopy diagnosis of precancerous cervical lesions. Cancer Med. colposcopy on the Caribbean coast of Nicaragua: a descriptive
2023; 12: 8690-99. mixed-methods study. Telemed Rep. 2021; 2: 264-72.
63. Chandran V, Sumithra M, Karthick A, George T, Deivakani M, 78. Ouh Y-T, Kim TJ, Ju W, Kim SW, Jeon S, Kim S-N, et al.
Elakkiya B, et al. Diagnosis of cervical cancer based on ensemble Development and validation of artificial intelligence-based analysis
deep learning network using colposcopy images. BioMed Res Int. software to support screening system of cervical intraepithelial
2021; 2021: 5584004. neoplasia. Sci Rep. 2024; 14: 1957.
64. Angara S, Guo P, Xue Z, Antani S. Semi-supervised learning for 79. Basu P, Sankaranarayanan R. Atlas of colposcopy – principles and
cervical precancer detection. In: 2021 IEEE 34th International practice. IARC CancerBase; 2017. Available from: https://screening.
Symposium on Computer-Based Medical Systems (CBMS). 2021. iarc.fr/atlascolpo.php.
p. 202-6. 80. Chen M, Xue P, Li Q, Shen Y, Ye Z, Wang H, et al. Enhancing
65. Ou Y, Xue Y, Yuan Y, Xu T, Pisztora V, Li J, et al. Semi-supervised colposcopy training using a widely accessible digital education tool
cervical dysplasia classification with learnable graph convolutional in China. Am J Obstet Gynecol. 2023; 229: 538. e1-e9.
Cancer Biol Med Vol 21, No 10 October 2024 879

81. Zhao J, Dai L, Zhang M, Yu F, Li M, Li H, et al. PGU-Net+: 86. Kirillov A, Mintun E, Ravi N, Mao H, Rolland C, Gustafson L, et al.
progressive growing of U-Net+ for automated cervical nuclei Segment anything. In: Proceedings of the IEEE/CVF International
segmentation. In: Proceedings of the First International Workshop Conference on Computer Vision. 2023. p. 4015-26.
on Multiscale Multimodal Medical Imaging: MMMI 2019, Held in 87. Xu H, Usuyama N, Bagga J, Zhang S, Rao R, Naumann T, et al.
Conjunction with MICCAI 2019, Shenzhen, China, October 13, A whole-slide foundation model for digital pathology from real-
2019. 2020. p. 51-58. world data. Nature. 2024; 630; 181-8.
82. Harangi B, Toth J, Bogacsovics G, Kupas D, Kovacs L, Hajdu A. 88. Hwang TJ, Kesselheim AS, Vokinger KN. Lifecycle regulation of
Cell detection on digitized Pap smear images using ensemble of artificial intelligence- and machine learning-based software devices
conventional image processing and deep learning techniques. in medicine. JAMA. 2019; 322: 2285-6.
In: 2019 11th International Symposium on Image and Signal 89. Niemiec E. Will the EU medical device regulation help to improve
Processing and Analysis (ISPA). 2019. p. 38-42. the safety and performance of medical AI devices? Digit Health.
83. Nanni L, Brahnam S, Ghidoni S, Lumini A. Toward a general- 2022; 8: 20552076221089079.
purpose heterogeneous ensemble for pattern classification. Comput 90. Gupta R, Kumar N, Bansal S, Singh S, Sood N, Gupta S. Artificial
Intell Neurosci. 2015; 2015: 909123. intelligence-driven digital cytology-based cervical cancer
84. Sarwar A, Sharma V, Gupta R. Hybrid ensemble learning technique screening: is the time ripe to adopt this disruptive technology in
for screening of cervical cancer using Papanicolaou smear image resource-constrained settings? A literature review. J Digit Imag.
analysis. Pers Med Univ. 2015; 4: 54-62. 2023; 36: 1643-52.
85. Promworn Y, Pattanasak S, Pintavirooj C, Piyawattanametha
W. Comparisons of Pap smear classification with deep learning
models. In: 2019 IEEE 14th International Conference on Nano/ Cite this article as: Wu T, Lucas E, Zhao F, Basu P, Qiao Y. Artificial intelligence
Micro Engineered and Molecular Systems (NEMS). 2019. strengthens cervical cancer screening – present and future. Cancer Biol Med.
p. 282-5. 2024; 21: 864-879. doi: 10.20892/j.issn.2095-3941.2024.0198