Evaluation of U-Net and Its Variants in Solving Upper Gastrointestinal Endoscopy Segmentation

Upper Gastrointestinal Endoscopy (or Upper GI Endoscopy) is one of the most commonly prescribed medical procedures to evaluate patients with problems in the upper GI tract, such as gastroesophageal reflux disease (GERD). An endoscopy will capture images of the digestive tract covering the esophagus, stomach, and duodenum for detailed analysis. During the analysis, the diagnostician needs to extract the necessary contours, surfaces or parts that are damaged, or anomalies from the image - This technique is called segmentation. Motivated by the fact that most publicly available datasets are specific to automated polyp detection while the abnormality areas caused by GERD disease can be varied like inflammation/ corrosion/ tearing on the images, the first goal of this work therefore is to collect and label an own dataset of upper GI Endoscopy images, especially those related to GERD. The dataset can be ready-shared for study, research community and non-profit development purposes. Secondly, to the best of our knowledge, it will be one of the first studies that evaluates U-Net and its variants in solving this particular problem. By conducting experimental detailed comparisons, we will analyze their results not only on segmentation accuracy but also on training time and average prediction time to highlight the most suitable model.