Study On Video Segmentation And Small Intestinal Protruding Lesion Detection Based On Wireless Segmental Endoscopy

Wireless capsule endoscopy(WCE)is a developed revolutionary technology with important clinical benefits.But the huge image data brings a heavy burden to the doctors for reviewing,locating and diagnosing the lesion images.Thus,in this paper,the research of WCE video is focused on two respects which are WCE video segmentation and lesion image detection.An efficient method in this paper helps clinicians segment the WCE video automatically according to stomach,small intestine,and large intestine regions.For efficient feature extraction,the color-saliency region detection(CSD)method is developed for extracting the potentially valid region of frame(VROF).To achieve better accuracy of boundary location,we apply Monitor-Judge model.Two novel features are designed for describing medical images: one is the color-texture fusion feature of visual perception(CTVP),constructed by Hue-Saturation histogram and grey level co-occurrence matrix(GLCM)features based on the maximum moments;another one is the color channels modelling of local binary pattern operator(CCLBP)that includes two kinds of patterns,grayscale pattern and color angle patterns.Moreover,support vector machine(SVM)classifier with proposed features is utilized to locate exact organic boundaries.Experimental results on abundant real WCE videos from China,Middle East and Europe,demonstrate that the proposed method improves the accuracy of the location of the organic boundary.After WCE video segmentation is implemented,two strategies are designed to help clinicians to detect small intestinal protruding lesion images automatically,which are the bag of local features(BOLF)and deep convolutional neural networks(CNN)based on transfer learning.On the one hand,we propose an improved bag of feature(BOF)method to detect protruding lesion images in WCE images.We extract different textural features(such as CCLBP)from the neighborhoods of the key points instead of utilizing the scale-invariant feature transform(SIFT)feature in the traditional BOF method.Specifically,we study the influence of different methods of computing key points,different local texture feature,the patch size and the number of visual words in terms of the classification performance.On the other hand,we also consider using deep Convolutional Neural Networks(CNN)with transfer learning for detect protruding lesion images about small intestine.The CNN approach requires a huge amount of data for a complex end-to-end classification process.However,in the fact,it is difficult to collect many lesion images.To overcome this issue,a transfer learning instance such that a pre-trained CNN can be employed as a later secondary process towards this application domain by explicit optimization.And this pre-trained CNN primarily is trained for generalized image classification tasks where sufficient training data exists.We empirically show that fine-tuned CNN features yield superior performance to conventional hand-crafted features on detect protruding lesion images when using current superior CNN architectures(such as InceptionV3).