Abstract
BACKGROUND: Capsule endoscopy is a non-invasive procedure used to visualize the small bowel, but optimal bowel preparation remains uncertain due to the absence of objective, automated evaluation techniques. Manual assessments by gastroenterologists are time-consuming and can suffer from inter- and intra-rater variability. This study aims to develop an objective, automated method to evaluate small bowel visualization quality in capsule endoscopy images. We developed a generative adversarial network to provide consistent assessments of image quality. MATERIALS AND METHODS: We randomly selected 1500 images from the Kvasir capsule endoscopy dataset, representing various disturbing factors like bubbles, debris, bile, and specular reflections, each with differing contamination levels. Clean and contaminated regions were annotated by three different annotators under the guidance of a gastroenterologist. A Pix2Pix network, known for its powerful image-to-image translation abilities, was employed for this task. It was trained on 300 annotated images and evaluated on a larger provided test dataset. RESULTS: The network achieved a Dice similarity score of 0.92 ± 0.07, an intersection over union of 0.86 ± 0.12, and an accuracy of 0.91 ± 0.09 on the test subset. CONCLUSIONS: The proposed model requires only a small set of annotated images during training and is well-suited for clinical applications, including quality control of capsule endoscopy videos, objective comparison of bowel preparation protocols, and assessment of antifoaming agents. Future work will explore applying the model to other modalities, such as colonoscopy and conventional endoscopy, and expanding its capabilities to segment other artifacts, enhancing its overall clinical utility.