An artificial intelligence system for human papillomavirus-positive dual stain cervical cancer screening reads p16-Ki67 slides

1. Lahrmann and colleagues developed an artificial intelligence (AI) system using dual stain triage for HPV-based cervical cancer screening to retrospectively analyze 9006 slides across 4 cohorts.

2. Results indicated that dual-stain reading can be reliably automated using this system

Evidence Rating Level: 2 (Good)

Study Rundown: Lahrmann and colleagues developed Cytoreader-Fusion, a model to automate p16/Ki67 dual stain reading (DS) for human papillomavirus (HPV) cervical cancer screenings. The advantage of DS testing is that it identifies patients at highest risk for cervical precancer more accurately than cytology alone. However, DS testing is limited by the labor-intensive nature of the process, and variation of interpretation between readers. The authors aimed to make this screening more accessible and usable by developing and externally validating Cytoreader-Fusion. Researchers locked the model prior to testing in December 2024. After this, they tested it blindly using 9,006 cervical and anal slides. The reference standard used was biopsy-confirmed Cervical Intraepithelial Neoplasia (CIN) (or Anal Intraepithelial Neoplasia [AIN]) grade 3+. In the independent cohort of 4,996 slides, the AI model’ssensitivity and specificity were 88.3% and 54.4%, respectively. Manual DS reading showed 78.1% sensitivity and 63.2% specificity, and Pap testing showed 74.5% sensitivity and 45.9% specificity. Cytoreader-Fusion could therefore be utilized in labs with limited or no prior DS experience. The strengths of the study included model locking before testing, having an independent team running the evaluation, and using biopsy-proven diseases as reference standards. Limitations included the retrospective design and the fact that the model’s specificity was lower than the manual reading in the independent cohort. Therefore, the AI model would screen more women without CIN 3+ as positive.

Click here to read the study in NEJM

Relevant Reading: Recommendations for Use of p16/Ki67 Dual Stain for Management of Individuals Testing Positive for Human Papillomavirus

In-Depth [retrospective cohort]: Model training used 1,919 slides, and testing used 9,006 slides across 4 cohorts. Kaiser Northern California contributed 3,094 slides (218 CIN 3+ cases), the NCI biopsy study contributed 653 slides (53 CIN 3+ cases), and Anal Cancer Screening Study contributed 263 slides. An independent external cohort contributed 4,996 slides. Both ThinPrep and SurePath slides were evaluated. As a reference standard, the authors used biopsy-confirmed CIN 3+ for cervical samples, and AIN 3+ for anal samples. Results were then compared to manual DS reading and Pap smear. In the independent cohort of 4,996 slides, the AI model showed 88.3% sensitivity (90% uncertainty interval [UI] 85.3 to 91.3) and 54.4% specificity (UI 52.9 to 55.9). Manual DS reading showed 78.1% sensitivity (UI 74.6 to 81.7) and 63.2% specificity (UI 61.7 to 64.7), while Pap testing had a sensitivity of 74.5% and a specificity of 45.9%.

Citation: Lahrmann B, Keil A, Miranda Ruiz F, Clarke MA, Egemen D, Grewal KK, et al. Closing the Automation Gap in HPV-Based Cervical Cancer Screening: Independent External Validation of an AI Model for Dual-Stain Triage. NEJM AI. 2026 Sep 24;3(10). doi: 10.1056/aioa2600270.

© 2026 2 Minute Medicine, Inc. All rights reserved. No part of this article may be reproduced, distributed or transmitted in any form or by any means without the prior written permission of 2 Minute Medicine, Inc. To license this content, contact 2 Minute Medicine, Inc. This article is for informational and educational purposes only; it is not medical advice and does not replace the judgment of a qualified healthcare professional.
Image: © 2 Minute Medicine