Point-of-Care Digital Cytology With Artificial Intelligence for Cervical Cancer Screening in a Resource-Limited Setting

This diagnostic study uses data from a single health center in Kenya to investigate the use of digital microscopy and artificial intelligence in a resource-limited area to detect abnormal cells in Papanicolaou tests.


Introduction
Inadequate access to microscopy diagnostics is a problem in limited-resource areas and impairs the diagnosis of common and treatable conditions. 1 Although significant advances have been made in digital microscopy diagnostics at the point of care (POC), their clinical implementation has been slow. 2 Here, we propose a digital diagnostics system in which microscopy slides are digitized at the POC and uploaded using local data networks for analysis with an artificial intelligence model based on deep learning. Cervical cancer remains a common and deadly cancer in areas without screening programs. 3 During the next decade, the disease incidence is expected to increase, and the yearly mortality is expected to double, with the largest burden of disease occurring in sub-Saharan Africa. 4 Ultimately, vaccinations against human papillomavirus (HPV) 5 have the potential to significantly reduce the disease incidence, but given that the full benefits of even the most efficient vaccination programs will take decades to be fully realized, millions of women remain at risk. 6 Therefore, screening tests remain essential, 7 and innovative POC diagnostic solutions are needed. 8 Conventional cytology screening (Papanicolaou test analysis) can drastically reduce the incidence and mortality of cervical cancer, but the manual analysis of samples is labor intensive, 9 is prone to variations in sensitivity and reproducibility, and requires medical experts to analyze the samples 10,11 ; this makes the process difficult to implement in resource-limited settings. 12 Human papillomavirus infections, which are the causative agent for cervical cancer, can be detected using polymerase chain reaction assays with high sensitivity and reproducibility. However, because most HPV infections are transient, the specificity for precancerous lesions is low. 13,14 In high-resource areas, both molecularand cytology-based screening methods are commonly used and are often combined (ie, cotesting) to improve the diagnostic accuracy. 15,16 Digital methods have been proposed to facilitate the visual analysis of Papanicolaou tests, but the development of fully automated systems has been challenging. 17,18 Although semiautomated systems for Papanicolaou test screening have been developed, 19 they are limited by the need for bulky, expensive laboratory equipment [20][21][22] and are not suitable for use at the POC or in resource-limited settings.
Recently, deep learning-based algorithms have been used for a large number of medical imageanalysis applications, with levels of performance even surpassing human experts in certain tasks. [23][24][25][26] However, studies on deep learning algorithms for analysis of cervical cytology smears have mainly analyzed only small areas of samples with instruments not suitable for POC usage. To our knowledge, no research has been conducted on the analysis of digital whole-slide images of entire Papanicolaou tests, captured in more challenging real-world clinical environments. [26][27][28][29] Thus, this technology has not yet been applied in basic laboratories that are able to perform simple staining procedures but lack access to molecular testing, where the need for improved diagnostics is highest. 28 In this study, we developed and implemented a novel POC digital diagnostic system at a rural clinic in Kenya, a country where cervical cancer is the leading cause of female cancer-related death. 30 Papanicolaou smears were collected at the clinic and digitized with a portable slide scanner, and whole-slide images were uploaded to a cloud platform using the local mobile data network for development and validation of a deep learning system (DLS). We measured the diagnostic accuracy for the detection of common forms of cervical squamous cell atypia with the DLS and validated the results by comparing them with the visual assessment of samples by independent pathologists.

Methods
that the patients understood the information provided. After this, signed consent from patients wishing to participate was obtained. Patients were compensated for travel expenses to the sampleacquisition site and informed of the test results, but they were not offered monetary compensation for study participation. This proof-of-concept diagnostic accuracy is reported in accordance with the Standards for Reporting of Diagnostic Accuracy (STARD) reporting guideline.

Study Design, Patient Cohort, and Collection of Samples
The research site for this study was a local clinic (Kinondo Kwetu Health Services Clinic, Kinondo, Kwale County) in rural Kenya (approximately 40 km south of Mombasa) (Figure 1). Papanicolaou smears were acquired from 740 women attending a regional HIV-control program (eFigure in the Supplement) between September 1, 2018, and September 30, 2019, from patient volunteers who fulfilled the inclusion criteria (nonpregnant, aged between 18 and 64 years, confirmed HIV positivity, and signed informed consent acquired) (eTable 1 in the Supplement). Eligible patients were assigned a study number, after which Papanicolaou tests were obtained from the patients by trained nurses and fixed and stained with the Papanicolaou staining method (eAppendix 1 in the Supplement). 31 After this, the staining quality was evaluated by light microscopy, after which the slides were digitized in the laboratory adjacent to the sample collection room at the research site. The slides were then stored in slide boxes and transported to the pathology laboratory (Coast Provincial General Hospital, Mombasa, Kenya). Patient records were stored digitally using the secured and password-protected web-based data-collection software REDCap (Vanderbilt University), running on a passwordprotected, encrypted local server in a locked room. Paper forms with patient data were stored in locked cabinets in a locked room at the clinic, accessible only to study personnel. Both digitized and physical slides were pseudonymized using study numbers, and no personal identifiers were uploaded to the cloud-based image-management platform. In cases of abnormal Papanicolaou tests, treatment expenses were covered by study funding, and treatment was arranged by a gynecologist (J.M.) in accordance with national guidelines. 32

Digitization of Slides at the Research Site
After the acquisition and staining of samples, Papanicolaou smears were digitized with a portable whole-slide microscope scanner (Grundium Ocus [Grundium]) ( Figure 1) and deployed in a laboratory space adjacent to the room at the local clinic where the samples were collected. The device features an 18-megapixel image sensor with a 20× objective (numerical aperture: 0.40) and captures images with a pixel size of 0.48 μm. The microscope scanner was connected to a laptop computer over a wireless local area network connection and operated via the web browser interface, Chrome (Google). The coarse focus for the scanner is adjusted manually, after which the built-in autofocus

4.
A, Study site location in Kenya. B, Slide processing, including staining bench and hood. C, Slide digitization equipment, including (1) laptop computer with access to the slidemanagement platform, (2) slide scanner, (3) mobile-network router, and (4) Papanicolaou test microscopy slide. routine is used for fine focus. Image files were saved on the local computer in tagged image file format and converted to the wavelet file format (Enhanced Compressed Wavelet [Hexagon Geospatial]) using a compression ratio (1:16) that was previously shown to preserve sufficient detail to not significantly alter the image-analysis results, 33

Development of a DLS for Detection of Cervical Cell Atypia
To develop a DLS for the detection of cervical cell atypia in the digitized Papanicolaou smears, we

Expert Visual Analysis of Samples
The analysis of physical slides was performed at the pathology laboratory at Coast Provincial General Hospital (Mombasa, Kenya) with light microscopy and performed by a trained pathologist (N.M.).
Slides classified as inadequate were excluded from the validation series (n = 29) (eFigure in the Supplement). Slides that were adequate for analysis (n = 361) were reviewed by the pathologist according to the Bethesda classification system. 34 For the analyses in this study, slides with findings recorded in the cytological report as LSIL or higher (ie, HSIL or higher) were included as slides with significant cervical cell atypia. The expert assessment of the digital slides was performed by remotely located, independent experts. For this process, all digital slides in the validation series were initially screened by a cytotechnologist with experience in cervical cytology screening, and digital slides with detected cellular atypia were reviewed by a pathologist with experience in Papanicolaou test analysis (L.K.). In accordance with generally accepted quality-control guidelines for cervical cytology screening, 35 10% of slides that were assessed as negative in this initial cytological screening were randomly selected and submitted for re-evaluation by the pathologist. The samples were reviewed by the 2 pathologists independently without access to results from the other pathologist or the DLS.

Statistical Analysis
General-purpose Stata statistical software, version 15.1 (StataCorp LLC) was used for analysis of the results. Statistical power calculations were performed with a sample-size formula, 36 assuming a

JAMA Network Open | Pathology and Laboratory Medicine
Digital Cytology for Cervical Cancer Screening in Resource-Limited Setting mean (SD) prevalence (P r ) of 8% (2%) for significant atypia in the study population 37 with an α level of .05 (and correspondingly Z 1−α/2 = 1.96) and a precision parameter (ε) of 0.10: where S N represents anticipated sensitivity; S P , anticipated specificity; and Z 1−α/2 , the standard normal deviate corresponding to α.
These calculations indicated a required target sample size of 304 for sensitivity and 19 for specificity with the assumed disease prevalence. All statistical tests were 2-sided unless otherwise stated, and the results were reported with 95% CIs. Evaluation of the performance of the algorithm was performed by calculating the area under the receiver operating characteristic curve (AUC) after plotting the measured true-positive rate (sensitivity) vs the false-positive rate (1 − specificity) for different thresholds of slide-level positivity. Interobserver agreement was measured using κ statistics.

Discussion
In this study, we implemented a POC digital diagnostics system at a peripheral clinic in Kenya and evaluated it for the analysis of Papanicolaou smears. The DLS achieved high accuracy for the detection of cervical squamous cell atypia, with AUCs of 0.94 to 0.96 and sensitivities of 96% to 100%, compared with the visual interpretation of digitized and physical slides. With the visual assessment of digitized slides as a reference, the number of false-negative assessments by the DLS was low, with 2 low-grade slides incorrectly classified as negative (although 4 high-grade slides were falsely classified as low grade). Compared with the visual analysis of the physical slides by the local pathologist, the DLS sensitivity was high for general atypia (100%) and high-grade atypia (100%) but low for low-grade atypia (21%), given that 11 of 14 physical slides that were assessed as low grade were classified as high grade by the DLS. The visual interpretation of Papanicolaou smears is known to be subjective, especially when assessing low-grade findings, 10,38 and accordingly, we observed variation between the experts' assessments of slides, with a lower threshold for the classification of findings as high grade by the pathologist who assessed the digitized slides. The DLS was trained with assistance from the experts who analyzed the digital slides, which possibly explains why the DLS classification showed higher agreement compared with these results. Notably, however, none of the slides that were classified as negative by the DLS were classified as atypical in the cytodiagnosis of the physical slides. Previous studies have reported encouraging results with the deep learning-based analysis of smaller cropped images from Papanicolaou smears 26,27,29,39 that were digitized with conventional slide scanners, but clinical application requires the examination of substantially larger sample areas. 28 In this study, we used routine samples collected at the clinic, and correspondingly, the whole-slide images were magnitudes larger than those previously analyzed, measuring on average 100 387 × 47 560 pixels; thus, the total number of pixels analyzed corresponded to approximately twice the number in the entire ImageNet database (>14 million images of everyday objects) at commonly used resolutions. 40 Papanicolaou smears may contain very limited numbers of isolated atypical cells, and robust algorithms are necessary to reliably detect such cells in these large and complex samples. In this study, we investigated the use of a DLS as a potential screening tool with a relatively low threshold for the classification of slides as atypical, to ensure high sensitivity at the potential expense of specificity; this method resulted in relatively high rates of false-positives for low-grade atypical slides. However, because this type of algorithm can operate using multiple configurations, sensitivity and specificity could be adjusted to match clinical requirements, with high sensitivity for screening purposes or higher specificity for confirmatory diagnostics. Importantly, our

Limitations
Because this is an early study, it has limitations. The DLS was benchmarked against 2 independent experts for the assessment of samples, but for the results to be directly comparable with other screening modalities, the ideal reference standard would be cervical biopsies with histologically confirmed precancers, which were not available here. Owing to the subjective nature of Papanicolaou smear cytology, this means that the results from both experts are not directly comparable with each other. Furthermore, even though the total number of slides collected was relatively large, the prevalence of slides with significant atypia was limited. Although these results are promising, increasing the amount of training data would likely improve the performance of the DLS and would be required before confirmatory diagnostic applications. Moreover, as this was a singlecenter study, the results might differ if the sample acquisition and preparation procedures are altered, and further work is needed to prospectively validate these results. Furthermore, because we evaluated only Papanicolaou smears from HIV-positive women, the results might differ owing to varying levels of prevalence in other populations (eAppendix 2 and eTable 3 in the Supplement).

Conclusions
In this diagnostic study, we developed a new system for deep learning-based digital microscopy at the POC, which was used for the analysis of cervical smears in cervical cancer screening. Results suggest that the detection of squamous cell atypia with the technology was feasible, with high sensitivity for slides demonstrating atypia, particularly for slides showing high-grade atypia. The clinical utilization of this technology could reduce the sample analysis workload for microscopists and provide a platform for general-purpose digital pathology, which is implementable in rural areas. As such, the technology here could create new opportunities to facilitate the diagnostics of a variety of diseases that are still underdiagnosed, especially in low-resource settings.