Title: Improving digestive organ classification from wireless capsule endoscopy images using deep learning
Authors: Supakorn Taweechainaruemitr; Padipon Thongjumruin; Nuttiwut Ektarawong; Kawee Numpacharoen; Amporn Atsawarungruangkit
Abstract: The location of a lesion is crucial information that Gastroenterologists must report using capsule endoscopy images. There have not been many studies that employ deep learning to automatically classify the location of the gastrointestinal tract. In this work, we created a deep learning model for identifying the organs of the gastrointestinal system (esophagus, stomach, small bowel and colon) using images from capsule endoscopy. The capsule endoscopies train set (670,051 images), validation set (411,702 images), and test set (216,978 images) are employ. The deep learning architecture is comprised of an InceptionResnetV2 Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM). On average, the accuracy is 92 percent, the precision is 89 percent, the recall (sensitivity) is 86 percent, the specificity is 96 percent, and the f1-score is 86 percent.