Pixel-wise annotation for clear and contaminated regions segmentation in wireless capsule endoscopy images: A...

Area of Science:

• Gastroenterology
• Medical Imaging
• Computer Vision

Background:

• Wireless capsule endoscopy (WCE) visualizes the small intestine but suffers from image quality issues due to contaminants like turbid fluids and air bubbles.
• These contaminants reduce mucosal view, increase review time, and risk missing pathologies, hindering accurate diagnosis and motility assessment.

Purpose of the Study:

• To construct the first multicentre, publicly available annotated dataset of WCE images to support the development of computer vision algorithms.
• To enable precise pixel-level segmentation for distinguishing clear and contaminated regions, including bubbles and turbid fluids, from normal tissue.

Main Methods:

• A dataset of 17,593 WCE images was curated from three databases (Kvasir, SEE-AI, CECleanliness).
• Images were annotated at the pixel level using ImageJ and ITK-SNAP software, creating binary and tri-colour ground truth masks.
• Annotations differentiated clear/contaminated regions and further classified bubbles, turbid fluids, and normal tissue.

Main Results:

• The dataset provides pixel-level annotations for 17,593 WCE images, distinguishing between clear and contaminated areas.
• It includes detailed segmentation of bubbles and turbid fluids, offering a valuable resource for AI model training.
• The multicentre nature ensures diverse representation of contamination patterns, enhancing model robustness and generalizability.

Conclusions:

• This curated dataset is the first of its kind, addressing the need for annotated WCE images to develop segmentation algorithms.
• It facilitates the creation of computer vision tools to identify contamination, improving diagnostic accuracy and efficiency in small intestine examination.
• The dataset enables robust model training and validation, paving the way for advanced WCE analysis in clinical practice.