EndoStyle: Improving AI for colonoscopy via endoscopic image style transfer

Artificial intelligence has shown great promise in detecting and classifying colorectal polyps during colonoscopy. However AI models are usually trained on images from many different endoscopic processors, while individual hospitals typically use just one. This mismatch can significantly reduce how well these systems perform in daily clinical practice.

To close this gap, we developed EndoStyle, a style transfer system that learns the visual characteristics of different endoscopic processors and converts images from one device to convincingly resemble those of another, without altering the underlying clinical content. In a multicenter study, experienced endoscopists rated real and converted images as equally realistic, confirming the system's credibility in a genuine clinical setting.

When EndoStyle was used to enrich the training data of a polyp detection AI, the results were striking: false positive rates dropped by more than 40% across two independent test datasets, meaning fewer unnecessary interventions and a more reliable tool for clinicians.

EndoStyle offers a practical and scalable path toward AI systems that perform consistently, regardless of which endoscopy hardware a center uses.

Publication

Joel Troya, Ioannis Kafetzis et al.
Gastrointestinal endoscopic image style transfer using EndoStyle to improve artificial intelligence prediction models
npj Digital Medicine 2026
Go to publication

This project was supported by the Eva Mayr-Stihl Stiftung and the Bavarian Cancer Research Center (BZKF).