Investigating cross-dataset abnormality detection in endoscopy with a weakly-supervised multiscale convolutional neural network

Abstract

The detection of abnormalities in endoscopic video frames can contribute in the early and more accurate detection of pathologic conditions. In this paper we present a novel Convolutional Neural Network (CNN) architecture for automatic detection of abnormal images in endoscopic video sequences. It features multiscale representation of the endoscopic images in its structure, and peephole connections contributing in enhanced generalization with less computational requirements. An important aspect of the proposed architecture is that it enables weakly-supervised learning, using only semantically annotated images. A novel cross-dataset experimental study is performed to investigate its generalization performance on various publicly available datasets. The results validate that the proposed architecture outperforms recent approaches, with results reaching up to 90.66% in terms of the area under the receiver operating characteristic. © 2018 IEEE.