Can AI cytology cut cervical-screening referrals without missing precancer? The strongest option also used HPV testing
A Wuhan study of 284,002 women found AI cytology alone referred fewer people than HPV testing but was less sensitive. Two combined pathways offered a better balance—within a retrospective, single-city analysis rather than a live randomised trial.
The short version
- What happened
- The retrospective study used records from 284,002 women aged 35–64 screened across 331 Wuhan primary-care facilities between July 2020 and August 2022.
- Who it affects
- For women, a lower referral burden can mean fewer invasive follow-up procedures and less anxiety, while missed high-grade lesions carry serious consequences. For screening staff, the relevant question is whether a combined pathway can preserve detection without exceeding local laboratory and colposcopy capacity.
- What remains unproven
- Retrospective population study
Keep in mind
Retrospective population study
One city in China
284,002 women aged 35–64
Nine hypothetical strategies
Histopathology unavailable for the full cohort
Verification bias corrected by imputation
Specific fixed AI system
No randomised implementation
No cost-effectiveness analysis
No mortality outcome
Source analysis · 1 linked source. Primary analysed source date: 30 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
AI cytology alone reduced referrals, but HPV testing found more precancer
The useful answer is not that one AI test won. In a peer-reviewed analysis of Wuhan's free community screening programme, AI-assisted liquid-based cytology referred 5.49% of women for colposcopy but detected an estimated 83.88% of cervical intraepithelial neoplasia grade 2 or worse (CIN2+). Pooled high-risk HPV testing detected 94.93% but referred 9.05%. That is a real service trade-off: fewer follow-up procedures can reduce anxiety and pressure on clinics, but lower sensitivity means more serious lesions may be missed.
The cohort included 284,002 women aged 35–64, with a median age of 48, screened at 331 primary-care facilities from 21 July 2020 to 13 August 2022. Every participant received both tests from one cervical sample. Researchers then evaluated nine predefined, hypothetical referral strategies on the same records. They estimated 1,727 CIN2+ cases after correcting for verification bias; 180 were invasive cancers.[1]
Combined pathways offered a stronger balance in this dataset
Two pathways stood out when HPV genotyping was available. Sending people with HPV16/18 directly to colposcopy and using AI cytology to triage the other high-risk HPV types produced 92.68% sensitivity and a 3.34% referral rate. A co-testing rule that referred people with higher-grade cytology, HPV16/18, or lower-grade cytology plus another high-risk HPV type reached 96.37% sensitivity and referred 3.70%. The authors' combined score of sensitivity and specificity was highest for that co-testing rule.
Those figures do not mean co-testing is automatically the best national policy. It requires both tests up front, and the paper did not perform a cost-effectiveness analysis. An alternative that referred anyone positive on either test reached 99.59% sensitivity but had lower specificity and substantially more referrals. Programmes must choose where they can safely place the balance between detection, laboratory capacity, follow-up access and unnecessary procedures.[1]
The distinction that matters
Four screening paths compared in the same Wuhan cohort
Corrected estimates for detecting CIN2+ show why referral burden and sensitivity must be read together.
AI cytology alone
83.88% sensitivity
5.49% referred
Pooled high-risk HPV
94.93% sensitivity
9.05% referred
HPV genotype then AI triage
92.68% sensitivity
3.34% referred
Specified co-testing rule
96.37% sensitivity
3.70% referred
The large cohort does not remove the study's main uncertainties
This was a retrospective comparison in one city, not a randomised prospective rollout of nine competing services. Histopathology was not available for everyone. The team linked electronic medical records and used Bayesian multiple imputation to correct verification bias, but disease status for unverified participants remained estimated rather than directly observed. Absolute accuracy could therefore shift, and infrastructure, HPV prevalence and follow-up behaviour may differ elsewhere.
The AI system was a fixed LANDING model trained on 188,542 cell-image patches from 8,329 women in nine Chinese provinces, with abnormal slides reviewed by cytotechnologists. That makes this evidence about a specific human-and-software pathway, not every AI cytology product. The work was funded by the Health Commission of Hubei Province; authors reported no competing interests. Prospective multi-region validation, direct implementation costs and outcomes beyond six-month records would change how confidently health services could act on these results.[1]