Investigation of Gas and Stool Region Extraction from Abdominal X-Ray Images Using U-Net with Semi-supervised Learning

Abstract

Chronic constipation presents with a variety of symptoms and conditions. Understanding the volume and location of gas and stool from X-ray images is crucial for selecting appropriate treatments. We propose an automatic extraction method for gas and stool regions from abdominal X-ray images using U-Net. However, annotating gas and stool regions in X-ray images is a time-consuming task even for experienced gastroenterologists, resulting in a limited amount of training data. Therefore, this study investigates the effectiveness of using semi-supervised learning with U-Net for extracting gas and stool regions from abdominal X-ray images. We conducted experiments under two conditions: 217 training samples annotated by gastroenterologists and an additional 49 pseudo-annotated samples created using the model, for a total of 266 cases. The extraction accuracy for the gas regions achieved DICE coefficients of 0.780 and 0.787, and for the stool regions, 0.647 and 0.664, respectively. In both regions, adding pseudo-(breakpoint)annotated training data improved the extraction accuracy, confirming the effectiveness of semi-(breakpoint)supervised learning in this method.