Study On Organ Classification Of Gastrointestinal Capsule Endoscope Images

Gastrointestinal(GI)tract diseases have been endangering people’s health for many years.Flexible endoscopy is entirely dependent on the clinician’s operation,which requires superb skills and rich practical experience.In recent years,wireless capsule endoscope(WCE)has gradually became a replacement of flexible endoscope in GI tract examination.There are huge amount of endoscope images recorded by WCE,as a result,experienced doctors need spend a lot of time and effort reviewing.And long-term review of WCE videos may lead to missed diagnosis and misdiagnosis.Therefore,in order to reduce the burden of medical staff,computer-aided system that can perform auxiliary diagnosis needs to be developed.The most basic and necessary step in this kind of system is the automatic organs classification of WCE images.The main content and contributions can be summarized as follows:(1)Wireless capsule endoscope video pre-processingFirstly,in view of the redundant images with high appearing probability in WCE video,a representative frame extraction algorithm based on inter-frame difference is proposed,where the multi-scale information of image is integrated through pyramid matching kernal model.Histogram intersection method is used to calculate the inter-frame difference of video.Similar frames are reduced by selecting frames with higher rate of inter-frame difference variation.Secondly,for the sake of reducing the effect of specular reflections on image feature extraction,the image inpainting algorithm based on adaptive window and dynamic search is studied,which is used to detect and inpaint specular reflections on representative frames.Thirdly,Aiming at the problem that the image segmentation based on pixel threshold is too fragmented,a detection method of shadows and highlights based on superpixel segmentation is proposed,which uses superpixel segmentation algorithm to pre-segment the image,and then automatically sets the threshold through the proposed self-tuning function to segment the shadows and highlights on the superpixel level.Moreover,a local contrast enhancement method is proposed to improve the contrast of the edge region and reduce the difficulty of parameter setting.(2)Organ classification based on feature fusionPresently,single feature or stacking several features is often uesed to train classifier in the field of WCE organ classification.But the performance is not satisfying because a single feature cannot characterize images comprehensively while stacking features is too dependent on quality of feature selection.Morevoer,lengthy feature vector requires dimension reduction algorithm to avoid over-fitting.Thus,a feature fusion framework based on discrimination power analysis for organ classification is proposed.Firstly,color and texture features are extract respectively.Then,the feature components with higher discrimination power are selected to form the fusion feature,which can not only retain the classification ability of each feature,but also reduce the feature dimension.The experiments show that the approach achieves in average 98.99% in classification accuracy on upper digestive tract organs and in average 98.50% on anatomical landmarks organs.Compared with the traditional method,this algorithm can improve the classification accuracy.(3)Organ classification based on bag of visual word(BOVW)Traditional machine learning methods for organ classification require powerful classifier which is time-comsuming task to achieve ideal result.The classification framework of BOVW can effectively deal with interference such as occlusion,illumination,rotation and change of view due to its encoding with local features.Moreover,the coding feature can achieve ideal classification accuracy only by using the linear classifier,which can greatly reduce the time cost compared with non-linear classifier.Thus,an organ classification method based on BOVW is proposed.And in view of the traditional BOVW without considering the sharing visual word among classes,a feature encoding algorithm and codebook learning method based on shared codebook are proposed to improve the performance of organ classification.The experiments show that the coding features generated by shared codebook have better classification accuracy compared with the codebook learning method based on locality-constrained linear coding.And the average accuracy of upper gastrointestinal tract organs is 99.65%,while anatomical landmarks organs classification accuracy is96.53% in average.