Related Experiment Video
Updated: Apr 30, 2026
09:21
A Novel Dual-Modal Deep Learning Approach for Real-Time Removal of Hepatic Fluorescence in Indocyanine Green-Guided Laparoscopic Cholecystectomy
Published on: April 17, 2026
31
Gastrointestinal endoscopic image style transfer using EndoStyle to improve artificial intelligence prediction
Joel Troya1,2, Ioannis Kafetzis1, Ronja Weber1
1Interventional and Experimental Endoscopy (InExEn), Internal Medicine II, University Hospital Würzburg, Würzburg, Germany.
NPJ Digital Medicine
|April 28, 2026
Summary
EndoStyle uses AI style transfer to make gastrointestinal endoscopy images compatible across different processors. This improves artificial intelligence (AI) polyp detection models, significantly reducing false positives in clinical settings.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Gastroenterology
Background:
- Artificial intelligence (AI) is vital for polyp detection in gastrointestinal endoscopy.
- Current AI models often struggle with images from diverse endoscopic processors, limiting clinical applicability.
- Processor-specific visual variations hinder the generalization of AI tools in real-world endoscopy.
Purpose of the Study:
- To develop and validate EndoStyle, an AI system for harmonizing endoscopic images across different video processors.
- To assess the visual and semantic fidelity of style-transferred images.
- To evaluate the impact of processor-specific AI model augmentation on polyp detection performance.
Main Methods:
- Developed EndoStyle, a StarGANv2-based style transfer system.
- Trained the system to mimic visual characteristics of five distinct endoscopic processors.
- Validated image fidelity using Fréchet Inception Distance and Learned Perceptual Image Patch Similarity.
- Assessed semantic similarity using three foundation models.
- Conducted a multicenter study with endoscopist evaluations.
- Augmented polyp detection model training with synthetic images.
Main Results:
- EndoStyle achieved high visual and perceptual similarity across different endoscopic processors.
- Converted images demonstrated strong semantic consistency with original content and style.
- Endoscopists rated real and synthetic images as comparably realistic.
- Augmenting AI training with synthetic images reduced false positives by over 40% in polyp detection.
Conclusions:
- EndoStyle effectively generalizes AI models across diverse endoscopic processors.
- The system offers a practical solution for improving AI performance in varied clinical environments.
- Synthetic image generation enhances the robustness and accuracy of AI-driven polyp detection.
