A novel AI-supported system for triaging HPV-positive women
Walter Prendiville, Maryluz Rol, Eric Lucas, Ankita Shastri,  JS Malliga, Sreeya Bose, Ranajit Mandal, Dipanwita Banerjee, Neerja Bhatla, Smita Joshi,  Richard Muwonge, Moleen Waison, Mike Chirenje , Bothwell Guzha Som Pathak, Partha Basu,
 
Background
HPV testing is gaining credibility as the optimum screening test for cervical precancer.   Management of HPV+ve women is challenging in many low or middle income countries. (LMIC). We have developed and investigated an inexpensive system of image capture and interpretation system to better manage HPV screen positive women, using standardised cervical images collected across diverse geographic regions.
Objectives
The objectives of this preliminary analysis were to (1) establish a baseline AI model using data from Asia, (2) assess model performance in an African setting, and (3) improve model performance and generalizability through retraining with multi-country data.
Methods
A portable, high-acuity cervical image-capture device (nGyn, fig 1) was developed with NSV (Allentown PA) and implemented during routine colposcopic assessment of HPV +ve women.  The image capture process was standardised.
1,820 women were recruited in centres in India and Thailand as part of the SAVECERVIX study.  Of these 687 had pathology reporting CIN2+.  A further 1200 cases were recruited in Zimbabwe as part of the EASTER study of which 62 had CIN2+.
Quality-control algorithms excluded poor-quality images. Images were ranked by focus and illumination, and the top-ranked images were retained for training. The images were then augmented and automatically cropped to the cervical transformation zone. The baseline model was developed by fine-tuning the ‘EfficientNet’ model.
A 5-fold cross-validation procedure selected a set of hyperparameters that improved the model's accuracy. To assess the discriminatory performance of a model with three distinct class outputs (normal/LSIL, HSIL, invasive cancer), ROC curves were calculated using test-set images. A comparison was made between the refined model and one trained on Zimbabwean data.
Results: A base model using acetic acid and/or Lugol's images was developed. The acetic acid model achieved 82.7% sensitivity and 73.9% specificity during training and had the highest discriminatory capacity between normal/LSIL and cancer (AUC 0.97) and the lowest between normal/LSIL and HSIL (AUC 0.74), compared to a model using Lugol’s iodine images.. This was then applied to images from Zimbabwe without pretraining to assess cross-setting performance. This revealed reduced accuracy. The model was then retrained using pooled data from India, Thailand, and Zimbabwe, resulting in 70% sensitivity and 81.1% specificity. Additional data collection continues inercontiinentally.
Conclusion
This ongoing study presents the early data from our use of a novel image capture and interpretation system to triage oHPV-positive women. The effectiveness of the system depends on capturing high quality images and the development of robust AI systems from representative datasets. The findings support the potential role of AI-based triage as a scalable, quality-assured solution to strengthen cervical cancer screening programs, particularly in LMICs. By addressing the key implementation challenges associated with HPV-based screening, this work contributes directly to global efforts aimed at eliminating cervical cancer.
 
 

nGyn system