AI matches human readers on HPV-positive cervical triage

Clinical takeaway: AI analysis performed on par with manual interpretation at identifying cervical precancer in HPV-screened patients. 

Primary human papillomavirus (HPV) testing made cervical screening easier to run at scale, but at the cost of specificity. This leaves clinicians with a pool of HPV-positive patients to sort before anyone is sent to colposcopy. Dual-stain cytology is a slide test that flags cells showing signs of HPV-driven change. It sorts those patients better than Papanicolaou (Pap) cytology and is an acceptable triage option in the US guidelines. Reading the slides is still manual, though, and depends on trained cytotechnologists, which limits how widely programs can offer the test. 

Automating that read is the obvious answer, but many pathology AI systems have been judged by how often they agree with human readers, which isn't the same as catching disease. An earlier AI system for dual-stain reading classified clear positives and negatives reliably, then wavered on borderline cells and needed separate models for the two liquid-based slide preparations. This latest retrospective validation tests a revised system against biopsy-confirmed precancer, the same standard applied to manual reading. 

The AI system, dubbed Cytoreader-Fusion, matched manual dual-stain reading for cervical precancer in the HPV-screened Kaiser Permanente cohort, with sensitivity of 91% against 89% and specificity of 51% versus 49%. In that cohort it also beat Pap cytology on both measures. In the colposcopy-referral Biopsy Study cohort, it ran close to manual readers on each, at 95% sensitivity versus 93% and 44% specificity against 45%. 

In the anal cancer screening cohort, the system held manual reading's 81% sensitivity and raised specificity from 33% to 47%, though the sensitivity range was wide, from 67% to 95%. In the final cohort, analyzed only after the system was locked, it caught more precancer than manual reading, with 88% sensitivity versus 78%. It gave up specificity to manual reading there, 54% against 63%, though it stayed ahead of Pap cytology on both measures. The uncertainty ranges around each pair did not overlap. A specificity-oriented setting, defined before testing, reached 62% specificity while keeping sensitivity at 86%, still above manual reading. 

This retrospective validation trained the system on 1,919 archived dual-stain slides and tested it on 9,006 more from four cohorts managed by the National Cancer Institute, including the Kaiser and Biopsy Study cohorts. A separate team at the institute held the biopsy results and ran every analysis. Manual dual-stain reading and Pap cytology were scored against the same biopsy-confirmed outcomes, with the system's standard setting as the primary comparison. 

The system has run in a health system's routine laboratory workflow, alongside manual reading and without changing any clinical report. Whether its calls change who gets referred to colposcopy hasn't been tested, which is why the authors call for prospective interventional studies with dedicated prospective confirmation of the anal result. 

"This study establishes a high-stringency validation framework for clinically deployable artificial intelligence in population-based cancer screening," the authors conclude. 

Source: Lahrmann B, et al. (2026 Sep 24) NEJM AI. Closing the Automation Gap in HPV-Based Cervical Cancer Screening: Independent External Validation of an AI Model for Dual-Stain Triage