Research On WCE Image Screening Based On Semantic Analysis

Wireless capsule endoscopy(WCE),as a novel technology used to record images of the patient’s digestive tract,has great significance for the diagnosis of gastrointestinal diseases.However,during the detection process,about 50-80,000 images containing a large number of disturbing images such as bubbles and impurities produced by each patient,and the physician’s job of diagnosing the disease by identifying the lesion images from these images is tedious and burdensome.Thus,many scholars have proposed computer-aided diagnosis methods based on artificial intelligence,but most of these methods ignore the impact of disturbing images such as bubbles and impurities of WCE images on disease diagnosis.Therefore,due to the problems of the large number and obvious semantics of WCE disturbing images,and traditional detection methods on disturbing images are poorly generic,this paper proposes a semantic analysis method to screen disturbing images from the WCE sequence.First of all,this paper proposes an algorithm based on a topical model for WCE disturbing image screening.The algorithm consists of three main parts: feature extraction,visual descriptor construction,and topic analysis.First,the local features of WCE images are obtained by traditional feature extraction methods such as HSV(Hue,Saturation,Value),Histogram of Oriented Gradient(HOG),Scale Invariant Feature Transform(SIFT),and Local Binary Patterns(LBP).Then,a large number of WCE image feature vectors are clustered by the K-means clustering algorithm to construct a set of visual descriptors that can express the content of the image.We obtain the visual descriptor distribution(also called the histogram of vocabulary distribution)of the images to achieve the semantic expression of the images.Finally,the topic analysis is performed on the lexical distribution histogram using the topic model to obtain the semantic classification of the images.The extensive experiments show our algorithm can effectively screen WCE disturbing images,while HSV and HOG methods can obtain a better detection rate and false detection rate but more missed detection.Secondly,we propose a feature extraction method based on an asymmetric convolutional auto-encoder for WCE disturbing image screening.Traditional features usually have poor generalizability and insufficient semantic information,leading to problems such as low recall and more missed detections in semantic analysis methods.In this paper,a WCE image screening algorithm based on asymmetric convolutional auto-encoder feature extraction is proposed to enhance image content representation.We investigate the effect of different parameters of asymmetric auto-encoder on disturbing image screening based on the semantic analysis method of WCE images based on the topic model.The extensive experiments show the asymmetric convolutional auto-encoder can effectively extract image features,resulting in better image screening performance.Also,the asymmetric convolutional auto-encoder improves the efficiency of the system compared with the symmetric structure of the auto-encoder.At last,this paper proposes a deep learning-based algorithm for WCE disturbing image screening.To address the problems of topic models relying on model initialization and large computation,we propose a WCE image semantic classification method based on classical convolutional neural networks.Meanwhile,we discuss the effects of different structures of asymmetric auto-encoder networks on semantic classification;and show the effective network structures in terms of screening performance and computational efficiency.The extensive experiments show that the classification network as a semantic judgment model can effectively improve the image classification accuracy,reduce false detections and missed detections,and obtain 97.30% screening accuracy,especially in difficult samples.In addition,the improved auto-encoder based on VGG-16 and Goog Le Net has a competitive advantage in WCE image feature extraction,which can effectively screen out bubbles and impurity images from the massive WCE images even by training on small data sets.