Artificial intelligence-driven clinical decision support system for cervical cancer detection in low- and middle-income settings: Technical development and deployment approach

Cervical cancer remains a major public health challenge, particularly in constrained-resource settings where financial, infrastructural, and human resources shortcomings undermine the effectiveness of screening programmes. While early detection and human papillomavirus (HPV) vaccination are recognized as effective preventive strategies, vaccine coverage remains insufficient in low-income countries, and several screening methods require costly equipment. In this context, Visual Inspection with Acetic Acid (VIA) has become a widely used alternative: low-cost and simple, it consists of applying a diluted acetic acid solution to the cervix, producing a transient whitening effect (acetowhitening). This phenomenon varies in both intensity and temporal dynamics depending on tissue type. Its assessment, however, relies on clinicians' interpretation, a process that is challenging even for experienced providers and thus remains inherently subjective. Artificial intelligence (AI) offers an opportunity to reduce the subjectivity of VIA interpretation by providing evidence-based clinical decision support. Yet most existing approaches rely on static images, ignoring the temporal dimension used in clinical practice. In addition, few account for benign conditions, such as inflammation, which can alter cervical appearance and complicate diagnosis. This thesis addresses these limitations through two components: 1. A technical component, focused on developing a smartphone-based clinical decision support system (CDSS) that leverages dynamic analysis of VIA; 2. A deployment component, centered on strategies to facilitate adoption by healthcare providers and ensure sustainable integration into clinical practice while reducing implementation risks. For the technical component, a dataset of VIA image sequences was collected in Cameroon and Switzerland to support algorithm development. These campaigns revealed image quality challenges, leading to targeted acquisition recommendations and the design of an automated quality control pipeline. The temporal evolution of acetowhitening was then analyzed using deep learning approaches to support diagnostic decision-making. For the deployment component, the work examined implementation in low-resource contexts, with Cameroon as a case study. Interviews with healthcare providers and patients revealed interest in using a CDSS for VIA, provided that data protection, workflow integration, and system transparency are ensured. Trust - between patient, provider, and technology - emerged as a central condition for adoption. The resulting system, smartCervix, is an mobile health (mHealth) solution that combines automated image quality control with temporal analysis of acetowhitening to support VIA interpretation.By combining a smartphone-based format, adaptability to constrained-resource contexts, and an emphasis on user trust, this CDSS emerges as a promising tool to strengthen cervical cancer screening. The thesis concludes by proposing a context-sensitive deployment approach for Cameroon, with potential adaptation to other low-resource settings.