AI-based staging addresses a critical question at the moment of colonoscopy: can lesion assessment become more accurate, more consistent, and more useful for treatment decisions in real time?
Within ECOPOP, this project focuses on the development of an AI optical staging tool designed to support endoscopists during colonoscopy. Its purpose is to improve differentiation between benign lesions, early colorectal cancer (T1), and more advanced disease (≥T2).
This matters because treatment decisions often begin with visual assessment. If a lesion is staged incorrectly during endoscopy, the next step in the patient pathway may also be wrong — whether that means underestimating disease extent or choosing a suboptimal removal technique.
Optical characterization of colorectal lesions remains challenging in routine and expert endoscopy alike. The appearance of a lesion may suggest one treatment strategy, while histopathology later reveals something more advanced.
When an advanced lesion is mistaken for an early one, treatment decisions may be inadequate from the start. This may lead to incomplete resection, inappropriate local treatment, delayed escalation, or avoidable burden later in the pathway.
The first endoscopic impression can shape whether a lesion is approached as benign, early cancer, or advanced disease — and therefore what kind of treatment follows.
If lesion extent is assessed incorrectly, the selected removal method may be insufficient and the overall treatment pathway may become more burdensome for the patient.
An AI tool may help reduce variability in visual staging by supporting endoscopists with image-based classification during colonoscopy.
More accurate staging may help align treatment intensity with real disease extent, which is central to ECOPOP’s broader clinical strategy.
The overall aim is not simply image recognition, but support for more accurate clinical decision-making at the point where treatment planning begins.
The tool is being developed using a combination of retrospective and prospective endoscopic datasets from multiple European centres. Histopathological assessment of resected specimens provides the reference standard against which the algorithm is trained and evaluated.
The project collects cases from clinical image and video databases, as well as from prospective ECOPOP trial-related pathways.
Expert endoscopists annotate lesion features, and these annotations are linked to histopathological ground truth to improve explainability.
The model is being developed to distinguish among benign lesions, T1 colorectal cancer, and ≥T2 disease in real time.
The project does not rely only on black-box prediction. It also aims to identify image features that can make the tool more interpretable to clinicians.
The development strategy combines multiple data streams to ensure variability and robustness across centres, imaging conditions, and lesion types.
A large retrospective image base includes approximately 2000 benign lesions and 150 colorectal cancer cases from Belgian centres led by KU Leuven.
Further retrospective cancer cases are contributed by hospitals in Norway and Italy.
Additional cases are expected from prospective pathways linked to ETHOS and T-REX, including lesions screened for trial eligibility.
Across sources, the project is expected to assemble a dataset of roughly 800 T1 and ≥T2 colorectal lesions, alongside benign comparator cases.
The goal is not just to create another classification model. The project is designed to support clinically relevant optical staging in a way that can eventually be used during real procedures.
It also places strong emphasis on explainability, feature-based interpretation, and real-world variability across centres — which makes the tool more relevant for eventual clinical use than a narrowly trained laboratory model.
The project is focused on developing an AI tool for real-time differentiation between benign, T1, and ≥T2 colorectal lesions.
The current structure includes data collection, image labelling, and algorithm development as the central work streams.
The concept begins from an already formulated technical foundation and is intended to move toward a functional proof-of-concept prototype.
Later project phases are expected to extend the work toward validation in more realistic clinical settings.
If successful, AI-based staging could help improve how colorectal lesions are assessed during colonoscopy more broadly — not only within ECOPOP trials, but also in future screening, diagnostic, and treatment pathways. In that sense, this project is both a technical development effort and a step toward more consistent, decision-relevant endoscopic assessment.
Questions about the project, collaboration opportunities, or its role within ECOPOP can be directed to the project coordination team.
Questions about the package, collaboration, or environmental endpoint integration within ECOPOP can be directed to the project coordination team.
Contact the team