Artificial Intelligence-driven Digital Cytology-based Cervical Cancer Screening: Is the Time Ripe to Adopt This Disruptive Technology in Resource-constrained Settings? A Literature Review

Abstract

Cervical cancer is still a public health scourge in the developing countries due to the lack of organized screening programs. Though liquid-based cytology methods improved the performance of cervical cytology, the interpretation still suffers from subjectivity. Artificial intelligence (AI) algorithms have offered objectivity leading to better sensitivity and specificity of cervical cancer screening. Whole slide imaging (WSI) that converts a glass slide to a virtual slide provides a new perspective to the application of AI, especially for cervical cytology. In the recent years, there have been a few studies employing various AI algorithms on WSI images of conventional or LBC smears and demonstrating differing sensitivity/specificity or accuracy at detection of abnormalities in cervical smears. Considering the interest in AI-based screening modalities, this well-timed review intends to summarize the progress in this field while highlighting the research gaps and providing future research directions.

Introduction

Cervical cancer continues to pose a major public health problem with over six million new cases and more than three million deaths worldwide [1]. Organized screening programs utilizing cervical cytology and high-risk HPV DNA-based testing have led to a significant reduction in both incidence and mortality in the developed countries [2]. However, the same could not be implemented in the majority of the low and middle-income countries due to the lack of infrastructural facilities and trained cytotechnologists and cytopathologists and the high costs involved [3]. Introduction of liquid-based cytology (LBC) improved cervical sample collection, preservation, smear preparation, and provided a smaller area to screen improving the screening efficacy. However, screening and interpretation errors still limit the performance of cervical cancer screening through cytology. Differentiation between categories such as ASC-US and LSIL or between reactive atypia and ASC-US/LSIL is largely subjective and constitutes most of the interpretation errors. HPV tests have been shown to possess higher sensitivity than cervical cytology along with providing objective results; however, these tests can not differentiate between a transient and persistent infection, the latter having a risk of developing cervical intraepithelial lesion. Hence, a positive HPV test necessitates inclusion of a triage (by cytology or visual inspection with acetic acid) before being referred for colposcopy [4, 5].

Automation has been devised to enhance objectivity of cervical smear interpretation through computer-assisted analysis, thereby reducing the screening and interpretation-related errors in cervical cytology. This may be undertaken as two approaches: capturing images of certain selected fields of the cervical smear that are analyzed using various algorithms or by presenting the complete slide for computer-assisted analysis. For the latter, whole slide imaging (WSI) converts the glass slide into a virtual slide that can then be analyzed using artificial intelligence (AI) algorithms. This approach has the potential of providing better sensitivity, specificity, and accuracy for detection of high-grade cervical intraepithelial lesions. There have been numerous studies involving the first approach of selected field analysis [6, 7]. However, the aspect of WSI images subjected to AI-based analysis has been the subject of relatively fewer studies. In the current era of primary HPV-based cervical cancer screening, AI-assisted cervical cytology has the potential of improving the accuracy of triage of HPV-positive cases prior to colposcopy. Similar enhancement of screening accuracy is also expected for cotesting approach using AI-assisted cervical cytology and HPV DNA detection. Hou et al. [8] and Allahqoli et al. [9] have recently reviewed the application of AI in cervical cancer screening—including cytology, HPV testing, and colposcopy. However, majority of the studies of cervical cytology included in these reviews were conducted using single cell images or selected fields from cervical smear.

This review aims to consolidate the progress in the field of artificial intelligence-enabled cervical cancer screening using whole slide scanned images (virtual slides) of cervical smear. The objective is to highlight the research gaps and limitations of the existing studies and provide future research directions on this aspect.

We conducted a literature search (PubMed, Cochrane Library, ProQuest, and Google Scholar) from 1990 till September 2022 using title/abstract search with terms “Artificial Intelligence” OR “Computer-assisted diagnosis” AND “whole slide imaging” OR “digital cytology” OR “whole slide scanning” OR “virtual slides” AND “cervical cancer screening” OR “cervical intraepithelial neoplasia” OR “high grade squamous intraepithelial lesions (HSIL)” OR “low grade squamous intraepithelial lesion (LSIL)” OR “Pap smear” OR “cervical cytology”. Research/original articles preferably published in English or those with available English translation and reporting on the use of artificial intelligence algorithms on whole slide images of cervical smears (conventional/ LBC) were included in this review. However, studies including only photomicrographs of selected fields of cervical smear or those that evaluated artificial intelligence algorithms on cervical biopsy/colposcopy images were excluded from the review. Similarly, the articles that described a method of segmentation of cervical smear images without an application of AI-based classification on whole slide scanned image were not considered. Research studies that employed only single cell datasets in their evaluation were excluded from the study.

Review

Automation in Cervical Smear Screening – Initial Attempts

The earliest attempts at automated screening systems for cervical cytology date back to 1950s when automated microscopes and analogue-based processing circuit software were used to analyze morphometric features. However, these systems were slow and cost-intensive and showed frequent overlap between the morphometric parameters of normal and atypical cells requiring manual cross checking.

Developments in the image processing techniques and computer display paved way for the introduction of the first commercial automated screening systems, PAPNET Testing System (Neuromedical Systems Inc., Suffern, New York, USA) and AutoPap 300QC system (Neopath Inc., Redmond, WA, USA) that received US-FDA approval for rescreening of manually screened conventional cervical smears in 1992 and 1995, respectively [10]. The introduction of LBC facilitated further the development of automated cervical cancer screening tools. Till date, two systems have received USFDA approval for use in primary cervical smear screening – BD FocalPoint™ GS Imaging System (Becton Dickinson, Burlington, NC, USA) in 1998 and ThinPrep Imaging System or TIS (Hologic Inc., Bedford, MD, USA) in 2003. Studies have demonstrated a sensitivity and specificity of 81.1–86.1% and 84.5%–85.1% for FocalPoint™ GS Imaging System-assisted cervical smear screening for neoplastic cells [11]. Similarly, the performance characteristics of TIS have been shown to be similar to manual screening (sensitivity 85.19%, specificity 96.67%) [12, 13]. However, these systems require specific methods of specimen collection, preprocessing, smear preparation, and imaging system that limit their utility in the resource-constrained countries.

Recently, two other systems have been developed recently – BestCyte^® Cell Sorter Imaging System (Cell Solutions LLC, Greensboro, NC, USA) from and Genius™ Digital Diagnostics System (Hologic Inc., Bedford, MD, USA). The BestCyte^® system incorporates remote access through a web-based software, cell annotation for peer review and review of the WSI by a cytotechnologist or pathologist [14]. A recently published study of 500 ThinPrep Pap test showed a significantly superior equivalency of the BestCyte screening compared to manual microscopy for positive smears (ASC-US + , LSIL + , and ASC-H +), suggesting an improved specificity leading to more efficient screening [15]. However, clinical validation of this system in a screening population is still required. The Hologic’s Genius™ Digital Diagnostics System, on the other hand, includes z-stacking during scanning along with improvements in the image analysis algorithms for expert review on a computer screen [14]. Hologic’s system has received CE-IVD for diagnostic use in Europe; however, their efficacy, sensitivity, and specificity are yet to be determined.

Artificial Intelligence, Machine Learning, and Deep Learning: Are They the Same or Different?

Availability and use of WSI and AI have opened new vistas for development of objective screening of cervical smears. AI is an umbrella term that encompasses machine learning (ML) and deep learning (DL). In ML, the process uses structured datasets of combination of various features selected by the programmer aside from giving weights to the features [16]. DL, on the other hand, automates the feature extraction part of the process, thus allowing for the use of larger datasets. Neural networks, a subfield of DL, are composed of input layer, variable numbers of hidden layers, and an output layer. There are several deep learning algorithms, of which a few have found utility in the field of cervical cancer screening.

• Convolutional neural networks (CNN) are also known as ConvNets. CNNs are mainly used for image processing and object detection. These networks include multiple layers like convolution layer, rectified linear unit, pooling layer, and fully connected layer. Instead of using the pixels of the whole image, a patch is used as an input to the consecutive layers that uses “convolution” to make a simplified feature map.

• Recurrent neural network (RNN) has connections that form directed cycles allowing outputs from long short-term memory networks to feed as input for the current phase, helping the network to memorize the previous inputs.

• Deep brief networks (DBN) consist of multiple layers of stochastic latent variables that have binary variables.

Among the deep learning network, two-stage object detectors such as FRCNN, R-FCN, and SPN as well as one-stage methods such as YOLO, single-shot multibox detector (SSD), and Retina Net have been tried for this purpose [17, 18]. Most of the studies included in this review utilized CNNs, some in combination with RNN, for the detection of abnormal cells in the cervical smears.

Some of the earlier works utilizing DL were conducted on cropped patches from WSIs or fields of view preselected by the pathologist. Development of a system that truly analyzes WSI images of cervical smears is a challenging task due to the large size of the images compared to the tiny size of the abnormal cells and the impact of the image acquisition process on the resolution of the final image. In this regard, conventional smear (CS) poses an additional challenge due to the area over which the smear is spread and the overlap of cells that is higher than that seen in LBC smears.

Studies Employing Convolutional Neural Network

Conventional Cervical Smears (Table 1)

Table 1.

Overview of studies included in the review

Author and year	Type of cervical smear included	AI techniques utilized	Number of smears included	Results (%)	Limitations
Studies utilizing convolution neural network (CNN) on conventional cervical smears
Holmström et al. (2021) [19]	Conventional 740 women living with HIV (18–64 years)	Deep CNN	Training and tuning set – 350 Validation set – 361	Sensitivity (HSIL) – 100 Specificity (HSIL) – 93.3 Sensitivity (LSIL) – 21.4 Specificity (LSIL) – 82.4 No false negatives	Histological confirmation of lesions not available Number of lesions small in the training set Interobserver variability in cervical cytology reading not accounted for
Wang et al. (2021) [20]	Conventional 143 cervical smears (retrospective)	CNN	Training – 97 abnormal cases Testing – 46 (8 normal, 38 abnormal)	Sensitivity – 89.8 Specificity – 100 Precision – 92.9	Small number of retrospective cases AI-based system can identify CIN2 + cells but not able to classify as CIN2, CIN3, or SCC
Studies utilizing convolution neural network (CNN) on liquid based cytology smears
Bhatt et al. (2021) [18]	Herlev (917 images) and SIPakMeD (4049 images) datasets 35 WSI for validation	Convolution neural network (CNN) with progressive resizing and transfer learning	Herlev dataset – 917 images SIPakMeD – 4049 images Validation – 35 WSI	Accuracy – 99.7 Specificity – 99.63	Small number of WSIs in the validation set
Lin et al. (2021) [21]	LBC 19,303 WSI	Deep CNN with dual-path encoder trained using synergistic grouping loss and rule-based risk stratification	Training – 13,486 WSI Validation – 2486 Testing – 3331	Sensitivity – 94.6 Specificity – 71.6	Built a large dataset of WSIs
Li et al. (2021) [22]	LBC	Deformable and global context aware region-based CNN-feature pyramid network	Training – 640 WSI (DHB dataset) Test – 160 WSI (80 positive, 80 negative)	Mean average precision increased by 6–9%	Low magnification of dataset used in the study Lack of exhaustive labeling of lesions No validation on prospective cases
Zhu et al. (2021) [23]	LBC	YOLOv3 for target detection, Xception and Patch-based models to boost target classification, U-net for nucleus segmentation classified using XGBoost and a logical decision tree	Training, testing, validation – 81,727 retrospective Clinical validation – 34,403 prospective	Sensitivity – 92 Specificity – 84.39	Lower accuracy for ASC-US lesions Difficulty in differentiating glandular atypia from adenocarcinoma LSIL, ASC-H, HSIL, and SCC combined as squamous lesions
Studies utilizing convolution neural network (CNN) with recurrent neural network (RNN)
Cheng et al. (2021) [24]	LBC 12 groups of datasets from five hospitals	Two consecutive CNNs to screen at low resolution (LR) and identify at high resolution (HR) followed by recurrent neural network (RNN) for WSI classification	Training set – 2315 slides Test set – 1270 slides	Sensitivity – 95.1 Specificity – 93.5	Used diverse slide staining and imaging modalities with good results
Kanavati et al. 2022 [25]	LBC Retrospective study 3073 WSIs	CNN and RNN	Training – 1605 Test – 1468 (in various sets)	Sensitivity – 62.4–88.6 Specificity – 61.9–92.0 Accuracy – 62.9–90.7	Sensitivity and specificity only modest in a clinical balance test set (95% NILM and 5% neoplastic)
Studies utilizing other deep learning algorithms
Tang et al. (2021) [27]	LBC	CNN-based object detection and classification projected using augmented reality (AR) technique	Training – 2167 Testing – 486	Sensitivity – enhanced from 86 to 95 Specificity – reduced from 95 to 90	Histological confirmation of abnormal cases not included
Zhao et al. (2016) [26]	LBC	Block-based classification algorithm and SVM classifier	Training – 1100 blocks including 100 suspicious blocks Testing – 960 blocks (12 images)	Sensitivity – 95 Specificity – 99.3 Accuracy – 98.98	H&E staining (in place of standard Papanicolaou stain) used in the study Relatively small number of images included
Bao et al. (2020) [28]	LBC Cohort study	Contour-based segmentation, deep learning algorithms	Training set – 8329 Validation set – 98,549 (manual and AI-assisted)	Sensitivity for CIN2 + 90.1 (84.3 with manual)	Unsatisfactory rate of AI-assisted cytology not calculated
Other applications of WSI and AI in cervical cytology
Wentzensen et al. (2021) [29]	Dual-stained smears	CNN with 4 layers and Inception-v3 with 48 layers	Training – 238 Validation – 3095	Cutoff of 2 cells Sensitivity – 88.1 Specificity – 61.5	–

Two studies developed CNN-based algorithms for the detection of abnormal cells in CS [19, 20]. Holmström et al. developed a point-of-care cloud-based deep learning system using a commercially available platform and demonstrated in a study among 740 women living with HIV. Their CNN-based system achieved a sensitivity and specificity of 100% and 93.3% for high-grade lesions while those for low-grade atypia were 21.4% and 82.4%, respectively. The system used in this study had the advantage of a cloud-based remote access to the WSI slides for expert assessment. An estimation of cost of the system (including slide preparation, digitalization of smears, and AI-based analysis) was approximately 3.70 USD per sample compared to the 18–27 USD being charged for cervical smear test per patient (with an additional 6.30 USD charged by the pathologist). However, the gold standard used in this study was cytological assessment by two experts rather than histopathological diagnosis and interobserver concordance of the two experts was not mentioned [19]. Also, the study included only women living with HIV where the prevalence of cervical cytological abnormalities is higher. The performance of this system in a general population is yet to be evaluated. Still, this system using a portable WSI scanner and DL algorithm raises hopes of an affordable AI-based system for cervical cancer screening in remote areas and field settings where trained pathologists may not be available.

The other study by Wang et al. included 143 deidentified conventional cervical smears from a tissue bank. Of these, 97 WSIs – ASC-H, HSIL, and SCC—were used for training while the test set (46 cases) included NILM, ASC-US and LSIL, ASC-H, HSIL, and SCC. The test set showed a precision of 93% and recall (true positive rate) of 90%, suggesting a potential for effectiveness in clinical use for detection of CIN2 + cells and allow cytotechnologists/cytopathologists to classify the cases appropriately [20]. However, certain issues such as difficulty in characterizing three-dimensional groups, smears with scattered single HSIL cells, and the inability to classify the smear as CIN2, CIN3, or SCC still need to be resolved in AI-based cervical cancer screening. The study by Wang et al. utilized hospital-based cases; validation of the system on a larger patient dataset preferably a screening population is still warranted [20].

LBC Smears

Most of the studies incorporating AI-based algorithms on WSI images have utilized LBC smears. In comparison to the conventional smears, LBC has the advantages of a smaller area to scan leading to reduced file size and scan time and lesser obscuration factors, especially cellular overlapping. Four studies developed and evaluated CNN-based algorithms for binary classification of LBC smears as normal and abnormal/atypical with variable sensitivity, specificity, and accuracy [18, 21–23].

The study from India by Bhatt et al. developed an algorithm using CNN with transfer learning and progressive resizing techniques for classification of WSI images into abnormal (dyskeratotic, koilocytotic), benign (metaplastic), and normal (parabasal, superficial-intermediate). The trained model achieved 99.7% accuracy, 99.7% precision, 99.6% specificity, and 99.7% recall (or sensitivity) when compared with expert analysis of WSI images by pathologists [18].

In another study, Lin et al. developed a hybrid model using a deep CNN with rule-based risk stratification (RRS). For this model, the epithelial lesions were grouped as epidermic (ASC-US, LSIL) and basal (ASC-H, HSIL, SCC). Of the 19,303 WSIs included in this study, 13,486 were used as the training set, 2486 as validation, and 3331 as testing set. Using an optimization for sensitivity and specificity, the RRS-based model displayed a sensitivity of 90.7% (96.8% for high-grade lesions) with a specificity of 80%. Attempts at improving the sensitivity to 94.6% overall led to a drop in specificity to 71.6% [21].

The network developed by Li et al. was based on faster region-based CNN (Faster RCNN)–feature pyramid network (FPN) architecture. The model was extended by introduction of deformable and global context aware (DGCA) RCNN. This model improved the detection of abnormal cells with a higher mean average precision that increased by 6–9% compared with other object detection models. However, the algorithm was not validated on a prospective set of cervical smears [22].

Zhu et al. developed an AI-assistive TBS (AIATBS), an integrated system comprising five AI models incorporating CNN and classifiers with logical decision tree models. The authors used 81,727 LBC samples for the training, testing, and validation of the system. Clinical validation was performed using 34,403 prospective samples. The validation set showed a sensitivity of 92% for detecting intraepithelial lesions (and categorizing as ASC-US, LSIL/ASCH/HSIL/SCC, AGC-NOS, and AGC-FN/AIS/ADC) with specificity of 84.39% [23]. The accuracy of categorization of cervical squamous lesions (ASCUS, LSIL, ASCH, HSIL, and SCC) ranged from 62.2% for ASC-H to 80.7% for ASC-US in the prospective samples [23].

Studies Employing Convolutional Neural Network with Recurrent Neural Network

Two studies employed CNN along with RNN to improvise upon the detection of abnormal cells in cervical smears. Cheng et al. combined CNN at low resolution (LR) for finding possible abnormal cells quickly and high resolution (HR) to confirm the abnormal cells. This was combined with a RNN for the calculation of the probability of a smear being positive. A small number of difficult-to-classify samples were included in the training of LR model to ensure sensitivity while a higher number of such samples were used for HR model to ensure precision. Validation of the system on an independent group of slides gave a sensitivity of 95.1% and a specificity of 93.5%. There were 0.8% false-negative (all were ASC-US) and 26.3% false-positive slides in the test set. A comparison of this screening system with the Hologic TIS showed a higher rate of true positives with the new system. Also, the new system could be applied to varied slide preparation, staining, and imaging techniques [24]. The authors suggested that their system could assist in reducing the cytotechnologists’ workload and prescreening of normal cervical smears.

Kanavati et al. also utilized a deep learning model consisting of a CNN followed by an RNN. An advantage of this model was the generation of a heatmap highlighting the areas with a high probability of neoplastic cells that could be further confirmed by the expert. They showed an accuracy of 88.5–90.7% with a sensitivity of 83.9–88.6% and a specificity of 90.4–92% [25].

Studies with Other Deep Learning Algorithms

Zhao et al. devised an algorithm whereby the LBC WSI images were divided into blocks (having 100 × 100 pixels) that were used for further computer-based processing using a support vector machine (SVM) classifier. The use of block-based analysis rather than cell segmentation reduced the computational complexity and enhanced the processing speed. Some of the texture and color features differed significantly between normal and suspicious blocks, and these were used as inputs for the SVM. A test set of 12 WSIs (960 blocks) showed an accuracy of 98.98% and a sensitivity of 95% for classification as normal/negative and abnormal/positive. However, the authors used H&E-stained smears rather than the standard pap-stained cases. The number of images used in the training and test set was also relatively small, and further validation is still required to assess the performance of this approach [26].

Tang et al. utilized WSI-based screening in an AI microscope with an augmented reality (AR) technique. For this, the authors utilized 2167 LBC smears for training the object-detection algorithm. The AR camera mounted on a microscope captured high-resolution images which were then transmitted to the AI algorithm in the attached computer. The binary output of the AI algorithm was then projected into the field of view through the AR method. The test set of 486 smears was interpreted by the same cytopathologists without and with AI assistance done 15 days apart. This system showed an improvement of sensitivity from 86.0% to 95.0% for detection of LSIL + cases with a slight drop in specificity from 95.0% to 90.0%. The category-specific analysis showed an enhancement of sensitivity from 83.7 to 92.3% for LSIL and 83% to 91.7% for HSIL using AI. One advantage of their system was that a traditional microscope could be plugged in with camera and AR display allowing real-time diagnostic help to improve cervical cancer screening. Hence, this could also find its application in cervical cytology training for cytotechnologist and pathologists and assist them to adapt to AI-based screening without major modifications to the routine working approach [27].

Bao et al. used an AI-assisted screening system in a population-based cervical cancer screening program. The AI system was trained on 8329 LBC images using a contour-based cell segmentation and classification of cells as normal and abnormal by the cytologists. In the validation phase, 98,549 cases were screened by manual reading as well as AI-assisted screening. An overall agreement of 94.7% was found between AI and manual reading and a 5.8% higher sensitivity for detection of CIN2 + by AI compared to manual reading with 0.4% lower specificity [28]. Hence, the authors suggested that AI-based population-level primary screening could be undertaken. However, the deep learning algorithms were not detailed in this study.

Other Applications of WSI and AI in Cervical Cancer Screening

An interesting application of cloud-based WSI platform was designed by Wentzensen et al. for evaluation of p16/Ki67 dual-stained smears for triaging the HPV-positive women. A cutoff of 2 dual-stained cells to call a cervical smear as positive gave a positivity rate of 42% and specificity of 61.5% for the detection of CIN3 + lesions by the algorithm while the accuracy of a cutoff of 100 cells approached that of HSIL cytology. A comparison of clinical efficiency showed that automated evaluation of dual stain had reduced colposcopy referrals with “most favorable ratio of colposcopies per CIN3 + detected.” This application of the WSI and AI platform needs to be investigated further for its utility in cervical cancer screening [29]. In this study, CIN3 + lesions were considered as the endpoint, though CIN2 lesions also possess a similar clinical connotation. Hence, another study using a CIN2 + endpoint is imperative to assess the clinical efficacy of this exciting approach.

Discussion and Summary

The present review focused on synthesis of findings from studies that utilized AI-based cervical cancer screening algorithms on the whole slide scanned images (WSI) of cervical smears to identify research gaps. This area is the current focus of intense research. Though the technology of AI-enabled cervical cancer screening using WSI images is very promising, the implementation of a successfully designed AI-based cervical cancer screening system may be met by challenges such as high cost of the scanner, requirement of extensive internal and external validation, and ethical guidelines.

The heavy cost of whole slide scanners may preclude their placement in remote areas undertaking screening. Solutions such as availability of low-cost portable scanners having adequate resolution that could be deployed in the field settings or transport of stained cervical smears to a nearby location possessing a whole slide scanner as a centralized facility that caters to multiple screening sites around it in a defined radius would need to be devised to augment the coverage of cervical cancer screening using this methodology.

Being a new technology, the use of WSI for diagnostic purposes needs guidelines. The College of American Pathologists (CAP) has provided guidelines for the validation of WSI systems to be used in diagnostic services [30]. The guidelines include three strong recommendations on the sample set size, concordance between virtual and glass slides, and washout period for examining the two sets of slides by the same pathologist. In addition, nine good practice statements have been provided which address important issues like appropriateness of the validation process, emulation of the real-world scenario, documentation, and other matters. However, cytology was considered to be beyond the scope these guidelines. A review by our group also highlighted the imperative need of regulations for the use of WSI in cytopathology [31]. A recently published systematic review of the validation studies of WSI in cytology showed that only one of the 25 included studies fulfilled all of the strong recommendations and good practice statements of the CAP guidelines [32]. With improvement in the scanning technology and incorporation of z-scanning to effectively scan cellular fragments, research interest has been growing for WSI in cytopathology. There is an imperative need of adapting the CAP guidelines for the field of cytopathology so as to provide guidance to the researchers and diagnosticians. Hence, future studies on the use of WSI in cervical cytology need to pay due attention to their study design and adhere to the current CAP guidelines in order to generate good evidence supporting or rejecting use of WSI in cytology.

Regulatory agencies such as the US-FDA, the Medical Device Regulation (EU), the International Telecommunication Union, and the National Medical Products Administration (China) have started framing regulatory requirements and policies for AI-based medical device use in healthcare. The US-FDA provides approval to such devices through one of the three pathways: premarket approval (for high-risk devices), de novo premarket review (low and moderate-risk devices), and premarket clearance or 510(k) [33]. The US-FDA has begun a review of its regulatory mechanisms for AI-based devices in 2019 and an update is expected soon. European countries, on the other hand, recognize the Conformité Européenne (CE) mark issued by accredited private Notified bodies, except for the lowest risk devices [34]. The AI-based medical devices are mandated to comply with the regulatory requirements and provide evidence of their reliability and repeatability of results. However, the major current challenges include data ownership, privacy, and protection, the black box nature of AI algorithms hindering with the understanding of how the algorithm analyzes and provides clinical decisions, safety evaluation in field settings, problems of bias or non-generalizability when the device is used in a new setting or population group, evaluation of the datasets used for training and validation of the device, and the ability of the regulatory bodies to keep abreast of the changes in the field of AI-based devices. Hence, the regulatory mechanisms in the field of AI-based systems for diagnostic or screening purposes need to be formed and strengthened to be able to tackle the challenges likely to be faced when these devices are put to clinical use.

We would like to mention a few limitations uncovered by this review. Very few studies were retrieved on the topic of use of AI on WSI of cervical smears, and these were published mostly in 2021 and 2022. This contrasts with the number of studies that have evaluated AI for classification of selected image frames of cervical smears. A recently published systematic review by Allahqoli et al. included several studies utilizing cervical smear images with accuracy ranging between 80 and 100% in differentiating cancerous and normal cells [9]. The use of WSI with AI in cervical cytology can preclude the first-level screening by cytotechnologists. In addition, this also provides an opportunity to healthcare providers doing cervical cancer screening in remote settings to seek opinion on the stained cervical smear from experts available elsewhere. These advantages formed the premise of the current review. Apart from being sparse in number, the reported studies showed variations in the type of input (conventional or LBC smears, different staining methods from Papanicolaou to H&E stain), differences in the scanning systems (magnification used for scanning, with or without z-scanning), and variances in the algorithms designed for AI-based analysis of the cervical epithelial cells. These differences have led to an irreconcilable variation in the performance characteristics of the systems designed in these studies. Another hindrance to the comparability of these studies is the lack/absence of gold standard (consensus opinion of two or more cytopathologists vs. histopathological diagnosis) as well as interobserver concordance. The available studies had differing populations such as general population or hospital-based women or high-risk populations such as women living with HIV or publicly available datasets. This variation also precluded their comparison with each other. We also found a lack of uniformity in the classification of cervical smears as binary or multiclass between the various studies.

To conclude, the use of AI-based algorithms for cervical cancer screening using whole slide imaging is in a nascent stage with a wide variety of research being undertaken by various groups. The current and future researchers in this field need to focus on collaborating with other groups to enhance the probability of translation of their results into clinical practice. At the same time, attention needs to be paid to development of suitable low-cost hardware for whole slide scanning and to support the AI-based algorithms that hold promise in real-world clinical scenarios. This would provide an impetus to all countries, developed or resource-constrained, to strive towards achieving the World Health Organization’s target for cervical cancer elimination.

Author Contribution

All authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by Ruchika Gupta, Neeta Kumar, Sompal Singh, Neelam Sood, and Sanjay Gupta. The first draft of the manuscript was written by Ruchika Gupta and Shivani Bansal, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Declarations

Ethics Approval

This is a literature review. Hence, ethics approval was not sought.

Consent to Participate

The study does not involve human participants. Consent to participate was not applicable.

Consent for Publication

The authors affirm that there were no human research participants and consent to publish does not apply.

Competing Interests

The authors declare no competing interests.