This article summarizes key observations from the peer-reviewed study “Point-of-Care Digital Cytology With Artificial Intelligence for Cervical Cancer Screening in a Resource-Limited Setting.”
Authors: Oscar Holmström, Nina Linder, Harrison Kaingu, Ngali Mbuuko, Jumaa Mbete, Felix Kinyua, Sara Törnquist, Martin Muinde, Leena Krogerus, Mikael Lundin, Vinod Diwan, Johan Lundin
Journal: JAMA Network Open, 2021;4(3):e211740
DOI: https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2777600
Despite advances in digital microscopy and cervical cancer screening, access to reliable diagnostic workflows remains uneven. Cervical cancer is highly preventable, yet it remains a major cause of death in settings where routine cervical cancer screening and other screening tests are limited. Conventional Pap test and pap smear analysis can help prevent cervical cancer by identifying abnormal cells, abnormal cervical cells, and other cell changes before disease progression, but cervical cytology depends on trained experts and laboratory capacity that may not be available in all settings.
This study examined whether digital microscopy supported by artificial intelligence could be implemented at the point of care in rural Kenya. The aim was to assess whether a cloud-based deep learning system could analyze Papanicolaou test slides and detect atypical cervical cells with high sensitivity. The authors note that HPV test methods can detect human papillomavirus and HPV infection with high sensitivity, while cytology remains important for specificity, and the two are often combined to improve diagnostic accuracy.
The study was conducted at a local clinic in rural Kenya among 740 HIV-positive women aged 18 to 64 years. Cervical smears were collected between September 2018 and September 2019 from participants attending a regional HIV-control program. After collection, the samples were fixed and stained using the Papanicolaou method, and staining quality was evaluated by light microscope review at the clinic lab.
Slides were digitized onsite using a portable whole-slide microscope scanner (Grundium Ocus) in a laboratory adjacent to the collection room. The digitized images were uploaded via local mobile networks to a cloud-based platform for remote analysis. This workflow combined specimen preparation, slide digitization, upload, and artificial intelligence–based assessment in a rural primary care setting, enabling analysis of full Pap smear whole-slide images rather than selected fields.
To evaluate AI in pathology, the researchers developed a deep learning system trained to detect low grade and high grade squamous intraepithelial lesions in digitized Papanicolaou test slides. In cytology terms, these findings correspond to clinically important atypia in cervical cancer screening and may reflect precancerous cells, cervical precancer, or other meaningful cell changes.
The model was trained on 350 whole-slide images and validated on 361 slides after exclusion of inadequate samples. More than 16,000 annotated regions were used, covering normal morphology and varying degrees of atypia, enabling analysis of large digital slides rather than selected microscopic fields. The analyzed whole-slide images were very large, reflecting full clinical samples rather than small image regions.
The study compared algorithm results with expert visual assessment of both digital slides and physical glass slides. Digital slides were first screened by a cytotechnologist and then reviewed by a pathologist when atypia was detected. Physical slides were independently reviewed by a local pathologist using light microscopy. This design allowed comparison with both digital-slide assessment and conventional glass-slide diagnosis.
After exclusion of inadequate samples, the study evaluated a validation set of digitized Papanicolaou test slides reviewed both by experts and a deep learning system. Digital-slide assessment identified cases of low-grade and high-grade atypia alongside a large proportion of negative samples, reflecting typical cervical cancer screening distributions.
The findings show that digital microscopy combined with artificial intelligence can support pap smear analysis with high sensitivity in a point-of-care setting. The deep learning system demonstrated strong agreement with expert review, particularly for high-grade atypia, while maintaining high negative predictive value for screening. The system was evaluated primarily as a screening tool, prioritizing sensitivity to ensure that atypical cases were not missed.
The authors note that visual interpretation of cervical cytology, especially for low-grade findings and atypical squamous cells, is subjective, which may help explain some of the variation in test result classification. This reinforces the role of digital pathology workflows not only in detection but also in supporting the evaluation of abnormal cervical cells across different reviewers and settings.
The practical importance of the study is that it demonstrates the feasibility of a point-of-care digital microscopy workflow for cervical cancer screening tests in a rural clinic. A system with high sensitivity and high negative predictive value may help identify slides with abnormal cells while reducing the number of normal slides that need full manual review.
This is particularly relevant in settings with limited pathology resources. If most slides can be screened digitally and only abnormal cases escalated, a large proportion of normal slides could be excluded from detailed review and clinician workload may be reduced through prioritization. In this way, AI-supported digital microscopy may contribute to more efficient pap smear analysis, help guide further testing, and improve access to timely treatment decisions in women’s health programs. The authors note that the platform may extend beyond cervical cancer to other diseases in low-resource settings.
The study has important limitations. It was an early proof-of-concept diagnostic study, and the reference standard was expert assessment of digital and physical slides rather than histologically confirmed biopsy findings. Because cervical cytology is subjective, differences between readers can affect direct comparisons.
In addition, the prevalence of significant atypia was limited, the study was conducted at a single center, and all participants were HIV-positive women. The results may therefore differ in other populations or where sample collection and preparation vary. The authors state that additional training data and further prospective validation would be needed before confirmatory diagnostic use.
This study demonstrates that digital microscopy with artificial intelligence can be implemented at the point of care for cervical cancer screening in a resource-limited setting. By combining slide digitization, cloud transfer, and deep learning–based analysis, the workflow enabled sensitive detection of atypical findings.
Within the limits of the study, the findings suggest that AI-supported digital cytology may help expand access to screening and support broader digital pathology use where expertise is limited.
A curated collection of digital pathology studies and references is available on Grundium’s website.
