IDDF2018-ABS-0253 A deep learning method for intestinal position locating in wce

Abstract

Background In a wireless capsule endoscopy (WCE) abnormal automatic detection system, the same image content may have different meanings in different locations. For example, visible blood vessels are usual in ileum but abnormal in the duodenum and upper jejunum, and bile is normal in the duodenum but indicating bile reflux in the stomach, as shown in figure 1. Therefore, the essential first step of building this system is to locate the intestinal position of the image. Methods To tackle this problem, a dense, connected convolution neural network (CNN) was adopted. The data were collected from several hospitals, consist of complete WCE videos of 16 patients, which has an average of 60 000 images per video. Those images were classified into four classes: oesophagus, stomach, small intestine and others by a gastroenterologist. The images before oesophagus part and after small intestine part were excluded because the large intestine disease is not the goal of the WCE. Results We used images of 10 videos (60%) for training and the rest 6 videos (40%) for testing. The performances are shown in table 1. We do not classify oesophagus vs. small intestine because of the order of WCE passing through the human gut. Due to the intestinal of patients are not clean enough, some images were fulfilled with food residue or faecal residue, which leads to the inability to determine its location. Rest of the misclassified samples are almost located at the junction of two parts because the intestinal wall of these places contains the characteristics of the front and back parts. Conclusions The CNNs show the great ability to distinguish the WCE images belongs to the oesophagus, stomach or small intestine. Some misclassified results were corrected based on the continuity of intestine for more robust performance, which will benefit the WCE abnormal automatic detection system behind.