Point-of-care digital cytology with artificial intelligence for cervical cancer screening at a peripheral clinic in Kenya

Cervical cancer is highly preventable but remains a common and deadly cancer in areas without screening programmes. Pap smear analysis is the most commonly used screening method but is labour-intensive, subjective and requires access to medical experts. We developed a diagnostic system in which microscopy samples are digitized at the point-of-care (POC) and analysed by a cloud-based deep-learning system (DLS) and evaluated the system for the detection of cervical cell atypia in Pap smears at a peripheral clinic in Kenya. A total of 740 conventional Pap smears were collected, digitized with a portable slide scanner and uploaded over mobile networks to a cloud server for training and validation of the system. In total, 16,133 manually-annotated image regions where used for training of the DLS. The DLS achieved a high average sensitivity (97.85%; 95% confidence interval (CI) 83.95-99.75%) and area under the curve (AUCs) (0.95) for the detection of cervical-cellular atypia, compared to the pathologist assessment of digital and physical slides. Specificity was higher for high-grade atypia (95.9%; 95% CI 94.9-97.6%) than for low-grade atypia (84.2%; 95% CI 79.9-87.9%). Negative predictive values were high (99.3-100%), and no samples classified as high grade by manual sample analysis had false-negative assessments by the DLS. The study shows that advanced digital microscopy diagnostics supported by machine learning algorithms is implementable in rural, resource-constrained areas, and can achieve a diagnostic accuracy close to the level of highly trained experts.