Objective Evaluation of Small Bowel Visualization Quality in Wireless Capsule Endoscopy Images Using Generative...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Capsule endoscopy visualizes the small bowel, but preparation quality assessment is subjective and lacks automation.
• Manual evaluation of capsule endoscopy images is time-consuming and prone to variability.
• Objective methods are needed to standardize and improve small bowel visualization quality assessment.

Purpose of the Study:

• To develop an automated, objective method for evaluating small bowel visualization quality in capsule endoscopy.
• To create a generative adversarial network for consistent image quality assessments.
• To overcome limitations of manual assessments in capsule endoscopy.

Main Methods:

• Utilized 1500 capsule endoscopy images from the Kvasir dataset with various contaminants.
• Annotated clean and contaminated regions across images with differing contamination levels.
• Employed a Pix2Pix generative adversarial network trained on 300 annotated images for image-to-image translation.

Main Results:

• The Pix2Pix network achieved a Dice similarity score of 0.92 ± 0.07.
• Achieved an intersection over union of 0.86 ± 0.12.
• Demonstrated high accuracy with a score of 0.91 ± 0.09 on the test dataset.

Conclusions:

• The developed model offers an objective and automated solution for capsule endoscopy quality control.
• It requires minimal annotated data, making it suitable for clinical applications.
• Future work includes application to other endoscopic modalities and artifact segmentation.