Convolutional Neural Network Based Small Bowel Lesion Detection in Capsule Endoscopy

Abstract

Although great advances in artificial intelligence (AI) for interpreting small bowel capsule endoscopy (SBCE) images have been made in recent years, its practical use is still limited. The aim of this study is to develop more practical convolutional neural network (CNN) algorithms for the automatic detection of various small bowel lesions.

SBCE images were collected for the training dataset from 526 SBCE videos taken at a single tertiary hospital. Abnormal images were classified into two categories: hemorrhagic lesions (red spot/angioectasia/active bleeding) and ulcerative lesions (erosion/ulcer/stricture). In this thesis, several approaches were attempted to improve the detection performance practically as follows: the preliminary study concerning class imbalance and problem definition, the sub-network fusion on decision-level to boost the sensitivity of the CNN model and lesion localization using a gradient class activation mapping (Grad-CAM), and the further data enrichment methods to SBCE images by human observation and a generative adversarial network (GAN).