Abstract:
Lesion location is a critical information that must be reported by
gastroenterologists when interpreting capsule endoscopy images. There are few
studies on how deep learning models can be applied to classify the locations of
gastrointestinal tract. Therefore, we aim to create a deep learning model for
classifying the locations of gastrointestinal tract (i.e., esophagus, stomach, small
bowel, and colon) based on capsule endoscopy images. Dataset of capsule
endoscopy images (n = 723,681 images from 174 patients) was divided into training
(n = 40,000) and testing (n = 683,674). The images were labeled into 4 organ
locations (esophagus, stomach, small bowel, and colon). We applied the deep
learning architecture (InceptionResnet V2) to the training dataset. Then, the
performance of the trained model was externally validated using the testing dataset.
The confusion matrix of the deep learning model, visualizing the performance of our
algorithm. The model classified capsule endoscopy images into 4 classes with
accuracy of 97.38%, precision of 89.25%, recall (sensitivity) of 94.07%, and f1-
score of 94.91%. The average prediction time was 20 milliseconds per image. In
addition, we have used ensemble for comparison and the results were slightly better
with accuracy of 94.21%, precision of 97.41%, recall (sensitivity) of 94.22%, and
f1-score of 95.39%. The deep learning model demonstrated an excellent
classification performance, which can be used as a building block for the ultimate
goal of creating a fully automated model for interpreting capsule endoscopy images.
