EndoStyle: Artificial Intelligence Image Transformation Tool for Colonoscopy

Overview

The study addresses the limitations of current AI systems in gastrointestinal endoscopy, which are tipically trained with data from a single type of endoscopy processor and have limited expert-annotated images. The investigators aim to develop and validate EndoStyle, an AI system that can generate images in the style of various processors from a single reference image. EndoStyle will be tested by showing endoscopists colonoscopy sequences with different image types to determine if they can distinguish AI-transformed images. Success would enhance AI training for diverse clinical setups.

The use of artificial intelligence (AI) in gastrointestinal endoscopy has become widespread. However, these systems are often only trained with data from a single type of endoscopy processor, which limits their applicability. In addition, the availability of images annotated by experts is limited, which affects data variability and thus the performance of AI systems. The aim of this study is to develop a new artificial intelligence (AI) based system (EndoStyle) and validate its authenticity by means of a survey among physicians, which is able to generate multiple images in the style of different processor types (including Olympus, Pentax and Storz) from a single endoscopy reference image. The investigators hypothesis is that the AI system is able to successfully change the image style of video processors, with the differences being imperceptible to the endoscopist's eye. The methodology consists of showing to multiple endoscopists 28 colonoscopy sequences of 10 seconds duration each. In each one of them 3 images will be shown that can be all the possible combinations of images belonging to positive control, negative control, and Endostyle (intervention group). By performing a statistical comparison of the percentages of selected images for each group the investigators will be able to establish whether the participants are able to distinguish the images transformed by the AI. If the results corroborate our hypothesis, our system could generate images that would allow a more customized training of AI systems for each clinical setup.

Conditions