Research On Portable Wireless Capsule Endoscopy Video Receiving System And Bleeding Pattern Detection In Endoscopic Images

Wireless Capsule Endoscopy (WCE) for the first time allows the painless endoscopic imaging of the entire small bowel, which enables physicians to directly view the Gastrointestinal (GI) mucosa to diagnose GI diseases, therefore the research about WCE is becoming more and more popular in clinical use at home and abroad .As a revolutionary technology, the existing WCE system cannot yet meet the clinical requirements. Especially when a huge number of pictures are generated in one inspection, it is very laborious and time-consuming for physicians to review the WCE images, which has limited the widely application of WCE.Supported by National Natural Science Fund of China (No. 60875061,31170968,30800235), 863 National High Technology Research and Development Program (No. 2007AA04Z234,2008AA04Z201,2006AA04Z368), Pre-research Project in Manned Spaceflight Field (No. 010203) and Shanghai Scientific Research Planning Project (No. 09DZ1907400), this thesis has carried out deep research on the portable capsule endoscopy video receiving systems and bleeding detection in WCE images. The principles of the portable capsule endoscopy video receiving system based on TMS320DM365 are explored, and the hardware and software design and architecture is discussed. On the other hand, two novel bleeding detection algorithms are also proposed.WCE can capture GI tract images within human body, and has the ability to code images into NTSC video with the frame rate of 30f/s. The WCE video is received by a RF receiver. Then the NTSC video is encoded into H.264 format digital video file, and saved in a SD card. After that, the system will decode the digital video and display it on a TFT-LCD screen. When the examination is completed, the video file can be reviewed on a computer by a doctor to make diagnosis.After analyzing the bleeding areas of WCE images, it is discovered that bleeding patterns manifest themselves by red colors different from colors of normal region. Intuitively, color appearance is also the first and most important basis for physicians to make diagnostic decisions. However, Light intensity in digestive tract is time-varying, and changes in the illumination can greatly affect the performance of bleeding detection if the descriptors used are not robust to these variations. Therefore, the illumination invariant color features are extracted, texture feature is combined with color features, and a Multi-layer Perceptron (MLP) classifier is utilized to classify WCE images. Experiments are designed to test the algorithms, and results show that the sensitivity of 98.6% and the specificity of 98.7% have been achieved. Compared to the state-of-the-art algorithms, the proposed algorithm is featured by high sensitivity.The bleeding detection algorithms based on MLP is good at solving the problem of WCE image classification; however, it has several disadvantages. MLP can be easily trapped into the local minimal, and has weak generalization ability. Besides, there are no guiding principles for the determination of the number of hidden layer. It also converges slowly. Therefore, algorithms based on Support Vector Machines (SVMs) are discussed here. SVMs is a kernel-based machine learning technique which has been widely used in real world classification problems in various domains. Due to its strong theoretical foundation, good generalization capability, and ability to find global classification solutions, SVMs is usually preferred by many researchers over other classification paradigms. This thesis incorporates spatial pyramids and photometric invariant color histograms, avoiding the loss of spatial information in traditional histograms. Then two kernel functions: histogram intersection kernel and chi-square kernel, which are especially suited for histograms, are utilized to construct SVM classifier. The algorithm is implemented using Matlab. Experiments show that the average best sensitivity and specificity reach 97.8% and 98.0% respectively. Compared with the previous algorithm based on MLP, this algorithm is stable. This has laid the foundation for the intelligent detections of other kinds of lesions.