Inclusive AI for Cancer Care: Model Adaptation and Human–AI Integration in Low-Resource Settings

Swiss partners

• Université de Lausanne: Yash Raj Shrestha

 MENA partners

• Arab Academy for Science and Technology: Noha Medhat Ghatwary

Presentation of the project

This project addresses a pressing challenge in global health and AI: how to adapt state-of-the-art (SOTA) models, originally trained on Western datasets, for equitable and trustworthy colorectal cancer detection in low-resource settings. Colorectal cancer is a leading cause of death worldwide, but early detection via colonoscopy can dramatically improve
outcomes. AI tools have shown strong potential to support clinicians during procedures; however, their performance and usability in diverse clinical contexts remain largely untested. This project fills that gap.

We pursue two main objectives. First, we assess whether existing SOTA AI models can accurately detect and classify polyp malignancy during colonoscopies in the Egyptian healthcare system. Second, we evaluate the fairness and generalizability of these models by testing and fine-tuning them using newly collected and annotated clinical data from Egyptian hospitals. The goal is to adapt these tools to local clinical realities while preserving diagnostic quality and reducing potential bias.

Beyond accuracy, we investigate how Egyptian endoscopists interact with AI in real clinical settings: How do they interpret recommendations? When do they trust or override AI outputs? What shapes usability and trust? To explore these questions, we conduct a mixed-methods study using surveys, interviews, and in-procedure observations. These
insights will inform best practices for integrating AI in medical workflows in resource-constrained contexts.

The project design prioritizes feasibility and impact. The Egyptian PI will lead clinical data collection and annotation across gastroenterology clinics, contributing substantial in-kind access and expertise. The Swiss team will lead the human–AI collaboration study, drawing on strengths in behavioral and organizational research. We also collaborate with Prof. Binod
Bhattarai (University of Aberdeen, UK) to benchmark model performance and support domain-adaptation through subcontracting.

Expected outcomes include: (1) a novel, validated colonoscopy image dataset from Egypt; (2) fine-tuned, context-adapted AI models; and (3) new evidence on human–AI interaction in low-resource healthcare. Together, these contributions promote inclusive, explainable, and clinically useful AI. The project builds on strong ties within the Global Gastrointestinal AI Network and is well-positioned for long-term collaboration and global impact.