AI algorithm development for cervical cancer screening in low resource settings

Cervical cancer remains a major global health challenge, with over 90% of cases and deaths occurring in low- and middle-income countries (LMICs). This disparity is primarily due to high prevalence of HPV infections in many LMICs, limited access to high-quality screening and diagnostic resources, as well as shortages of trained personnel. Conventional screening methods, including Visual Inspection with Acetic acid (VIA) and colposcopy, heavily rely on visual assessment, which introduces subjectivity and variability, especially when conducted by less-experienced clinicians. While colposcopy has been standard for detecting cervical intraepithelial neoplasia (CIN), it is resource-intensive and demands trained personnel, making it inaccessible in many low-resource settings. AI-based Automated Visual Evaluation (AVE) models offer a promising solution, providing consistent, reproducible classifications that can improve diagnostic reliability. Conventionally, AVE algorithms have used a single image typically acquired around 1-2 minutes after the application of acetic acid. In this proposal, we aim to develop advanced AI risk models for cervical cancer screening, using timeseries imaging to capture acetowhitening patterns after acetic acid application on the cervix acquired on a low-cost device. The model will leverage self-supervised learning (SSL) for precise image registration across time points and state-of-the-art networks for classification. This approach enhances the temporal alignment of images, crucial for capturing lesion dynamics. Our model will offer a comprehensive risk stratification that improves diagnostic accuracy by combining imaging data with HPV genotyping and AVE. We will evaluate the model’s performance using datasets representing high- and low-resource devices and settings. Success in this project will enable consistent, objective, and scalable cervical cancer screening, with potential impact in low- and middle-income countries, aligning with WHO’s goals for cervical cancer elimination. Additionally, we will develop and validate novel algorithms for the analysis of time-series datasets that can be applied to many image analysis problems in oncology.