Deep Model-Based Semi-Supervised Learning Way for Outlier Detection in Wireless Capsule Endoscopy Images

Abstract

Wireless capsule endoscopy (WCE) has become an irreplaceable tool for diagnosing small intestinal diseases, and detecting the outliers in WCE images automatically remains as a hot research topic. Considering the difficulties in obtaining sufficient labeled WCE data, it is necessary to develop the diagnosis model which works well with only little labeled or even unlabeled training samples. In this paper, a novel semi-supervised deep-structured framework is introduced to solve the problem of outlier detection in WCE images. The key idea of our model is to mine the anomalous graphical patterns existed in the image by analyzing the spatial-scale trends of sequential image regions. Three main contributions are concluded: 1) we integrate a convolutional neural network into long short term memory network, so that the intrinsic differences between outliers and normal instances could be captured. Besides, 2) a assessment model is built by using various signs of anomaly occurrence and fake outliers knowledge learned during the training stage, which enhances the outlier alarm accuracy significantly. Furthermore, 3) a nest-structured training method is proposed, which helps our model achieving efficient training process. Experimental results on the real WCE images demonstrate the effectiveness of our model.