2024 年 17 巻 p. 18-26
This study explores the comparative accuracy of two state-of-the-art algorithms, YOLOv3 and faster region-based convolutional neural network (R-CNN), in detecting endoscopy artifacts using the EAD2019 dataset. YOLOv3, primarily used for real-time tasks, and Faster R-CNN, which employs a two-step object detection process, exhibit variable performance based on the object characteristics. The analysis performed in this study focuses on identifying the objects or classes where each algorithm performs better. We conduct experiments to support our findings. We introduce a novel metric that quantifies the difference in average pixel intensities inside and outside the bounding boxes of detected objects. This metric forms the basis of a proposed ensemble method, allowing the method to effectively utilize either YOLOv3 or Faster R-CNN, depending on the characteristics of each class. The proposed method demonstrates an improved average precision score compared to using either algorithm separately. This research provides valuable insights into object detection in endoscopy, potentially enhancing artifact detection accuracy in medical imaging.