Deep Learning-based Landmark Identification for the Upper Gastrointestinal Track in Endoscopic Images

Abstract

Accurate identification of landmarks is critical for effective diagnosis and treatment in endoscopy, particularly in the upper gastrointestinal tract. However, there are many similar structures inside the stomach, and it might be difficult to accurately locate landmarks in camera images because of other factors such as air bubbles and the narrow field of view of wired endoscopic images. This study presents a comparative analysis experiment conducted with a model that can identify anatomical landmarks of the upper gastrointestinal tract with high accuracy through small-scale data-based local augmentation. We used five classes captured by esophagogastroduodenoscopy criterion, preprocessed medical image data to address the class imbalance, and compared the accuracies of ResNet50, MobileNetV2, and DensNet265 models. We used a dataset comprising 2,546 images of patients who underwent upper gastrointestinal endoscopy at Yonsei Severance Hospital. We augmented 4,632 images and evenly distributed them across five classes. Our results indicate that this is the most accurate model for improving diagnosis and treatment in upper gastrointestinal endoscopy. The ReseNet50 model achieved the highest accuracy at 74.88%, followed by the MobileNetV2 model at 78.91% and DensNet265 at 84.70%.