AI-Assisted Cytology Meets HPV Genotyping in a 284,002-Woman Screening Cohort
Cervical cancer screening is often judged on a narrow balance: detect as many cervical intraepithelial neoplasia grade 2 or worse lesions as possible while avoiding an unmanageable flood of colposcopy referrals. In a large community-based screening cohort in Wuhan, China, investigators evaluated how artificial intelligence-assisted liquid-based cytology and high-risk human papillomavirus testing perform alone and in combination across nine hypothesized strategies. The scale matters. With 284,002 women included, this is the kind of population in which small differences in sensitivity or referral burden become operationally meaningful for public health systems.
The central signal is straightforward. After correcting for verification bias, HR-HPV testing alone outperformed AI-LBC alone on sensitivity, reaching 94.93% versus 83.88%, and posted a higher Youden’s index, 0.864 versus 0.789. The cost of that gain was a higher colposcopy referral rate, 9.05% versus 5.49%. In screening programs, this trade-off is not abstract. Every extra percentage point in referral rate translates into more women needing follow-up procedures, more clinic capacity consumed, and more opportunities for attrition between the screening result and diagnostic evaluation.
Referral Burden Rose as Sensitivity Improved
The performance pattern reflects the biology of cervical carcinogenesis. HR-HPV testing captures the viral driver upstream, before morphologic abnormality necessarily becomes obvious under cytologic review, which helps explain its stronger sensitivity. Cytology, even when augmented by artificial intelligence, depends on visible cellular change and therefore can miss earlier lesions that are biologically active but not yet conspicuous. AI can tighten interpretation and reduce human variability, yet it still operates within the limits of what is present on the slide.
Among the triage-based strategies, one of the more efficient options was AI-LBC with reflex HR-HPV. Strategy 3 delivered a sensitivity of 82.90%, a referral rate of 2.67%, and a Youden’s index of 0.807. That profile makes it attractive where minimizing downstream procedures is a priority, especially in systems with finite colposcopy capacity. The lower referral burden is not a minor administrative detail; it directly shapes whether a screening program can be sustained at scale without saturating diagnostic services.
A different triage pathway, HR-HPV genotyping with reflex AI-LBC, pushed performance higher. Strategy 7 reached 92.68% sensitivity, a Youden’s index of 0.899, and a referral rate of 3.34%. The data suggest a practical sequence: use viral genotyping to identify women at higher biological risk, then bring AI-assisted cytology into the workflow to refine who moves on to colposcopy. In effect, the genotyping result provides the first filter, and the cytology result sharpens the second one.
Co-Testing Delivered the Highest Overall Accuracy
The strongest overall balance came from co-testing with AI-LBC and HR-HPV genotyping. Strategy 9 achieved the highest Youden’s index at 0.932, alongside a sensitivity of 96.37% and a referral rate of 3.70%. The number is important because Youden’s index captures the combined behavior of sensitivity and specificity in a single measure, making it useful when the goal is to compare screening strategies that impose very different diagnostic burdens.
In practical terms, co-testing maximizes detection while keeping referrals well below the level seen with HPV testing alone. That is a clinically meaningful compromise for organized screening systems that can support dual testing. It also gives a glimpse of where artificial intelligence may fit best: not necessarily as a replacement for established molecular testing, but as a force multiplier that improves triage, reduces ambiguity, and helps allocate follow-up resources more intelligently.
The findings also argue against treating AI-assisted cytology as a single fixed use case. AI-LBC performed well as a standalone screening method, as a triage tool, and as a component of co-testing. Its value changes depending on the surrounding infrastructure. In settings where molecular testing is limited, AI-assisted cytology may still function as a workable primary screen. Where genotyping is available, pairing the two modalities creates a stronger diagnostic net.
Why the Balance Between Detection and Capacity Still Dominates Screening Design
Cervical cancer prevention depends on two linked processes: identifying women with meaningful precancerous disease and ensuring they actually reach colposcopy. A highly sensitive strategy that overwhelms referral pathways can backfire if the health system cannot absorb the downstream load. A more selective approach may be easier to run, but missed disease is the price when sensitivity falls too far. These results map that tension with unusual clarity.
For program planners, the key question is not whether one strategy is universally superior. It is how much sensitivity a given system can afford to trade for lower referral burden. In a citywide screening context like Wuhan, the answer may differ from what would work in a smaller clinic network or a resource-constrained region. The data support a tiered view: Strategy 7 and Strategy 9 emerge as the leading options when HR-HPV genotyping is available, and the final choice depends on whether the priority is maximal detection or tighter control of referrals.
The study also adds a useful real-world dimension to the debate over AI in cervical screening. Artificial intelligence is not replacing the biological signal; it is helping interpret it. The strongest performances came when AI-assisted cytology was embedded in a broader testing sequence rather than asked to shoulder the entire screening burden alone. That may prove to be the more durable model for implementation: AI as an adjunct that improves workflow efficiency and diagnostic precision, not as an isolated answer to cervical cancer prevention.
Even so, the central uncertainty remains operational, not conceptual. Screening policies must still decide how to deploy these strategies across populations with different laboratory capacity, referral infrastructure, and acceptable thresholds for over-referral. The current data provide a strong comparative framework, but the best choice in practice will hinge on local program design and the resources available to carry women from screening to diagnosis without delay.