Feature-Based Distributed Learning for Colorectal Cancer Detection from Colonoscopy Videos

Clinical Need and Research Gap

Despite its effectiveness, colonoscopy still misses many flat and sessile polyps due to human limitations and image variability. Current AI tools rely on large, shared datasets, which are often unavailable due to privacy laws like GDPR. Moreover, hospital imaging hardware varies widely, making it difficult for centralized models to generalize. There is an urgent need for adaptive, privacy-compliant AI solutions that can work with diverse video sources.

Objectives and Methodology

This research lays the foundation aims to bridge the methodological innovation and clinical impact. It introduces a distributed, feature-based learning framework that uses robust optimization to handle rare and heterogeneous data in decentralized settings. Clinically, it offers a privacy-preserving approach that helps hospitals improve CRC detection without sharing raw data. Using the FIFA algorithm, it enables adaptable, interpretable AI across diverse imaging environments.

Impact and Future Applications

The results will serve as a stepping stone toward real-time AI in endoscopy, offering an approach that adapts to colonoscopy imaging environments and supports future clinical integration. Continued development of this project could reduce miss rates and enable earlier CRC detection—especially for subtle lesions—without altering hospital hardware or workflows. With survival rates over 90% at Stage I, this may save thousands of lives and billions of euros in healthcare costs.