Design A Dual-camera Capsule Endoscopy Based On Wireless Energy Transmisson System And Bleeding Image Recognition

As a safe and convenient, painless and non-invasive, the whole digestive tract detection tool, Wireless capsule endoscopy(WCE) can get the most intuitive information of the lesion by gathering internal images of the human digestive tract. WCE has a unique advantage especially in the human small intestine where other traditional plug-in endoscopy is can hardly reach. In recent years, WCE research has become one of the hotspots in the field of medical devices. WCE greatly reduces the pain of the patient, however, at the same time the transmission of the image frame rate, image quality and running time still can not fully meet the requirements of clinical use. Especially the huge number of collected images of digestive tract is up to tens of thousands, which require the medical staff to check one by one bringing medical staff a heavy burden.This problem not only reduces the efficiency of the medical staff but also hinders the practice and use of WCE.This paper relied on the National Natural Science Foundation(No. 31170968) and Biomedical Engineering, Shanghai Jiaotong University Cross Foundation. This paper conducts a deep study of the dual-camera video capsule endoscopy systems based on wireless energy transmission and bleeding image recognition algorithm.We design a new type of dual-camera WCE system based on OV6920, including hardware and software framework design, mechanical design and a new design of three-dimensional receiving coil.We also focus on the bleeding image recognition algorithm, including the contrast and analysis of color spaces, bleeding feature extraction method comparison and parameter optimization of SVM and so on.Currently, the battery for WCE can only maintain the effective working time for about 6-8 hours which is less than a complete digestion cycle time.So it will produce some detecting blind spots. In this paper, we have a deep research of WCE system based on wireless energy transmission. First of all, the system is an innovative design.We upgrade the traditional single-head to dual-camera and the frame rate can reach 30 f / s. The load frequency of two transmission channels are respectively at 991.4MHz and 1.013 GHz with 21.6MHz difference, which keep the images be received clearly; In addition, in order to further reduce the size of the WCE system, we design a new three-dimensional wireless receiver coil which contains a hollow cylinder core made with high permeability Mn-Zn material and a new oval receiver coil. The advantage is that part of the hardware circuit can be placed in the hollow magnetic loop to reduce the length of the entire system. The experiment proves that new coil system is in a stable operation and the system reduces the length by 6mm.In addition to dual-camera WCE system, this paper also explores the statistical learning methods and the use of support vector machine as capsule endoscopy bleeding recognition classifier. SVM has the advantage of small samples classification, good arithmetic convergence speed, and strong generalization ability and so on. This paper explores the principles of RGB, HSI, HSV color spaces and the conversion method among them. This paper studies three common feature extraction bleeding ideas: color histogram-based global image feature extraction method, local image-based feature extraction method and pixel-based feature extraction method. This paper presents a bleeding characteristic parameter selection method, which combines the pixel-based feature extraction method and the local image-based feature extraction methods. Firstly, three independent RGB vector are sampled with a same interval, reducing the range of vector. Secondly, select the feature vector of red pixels over the sum of the red and blue pixels feature vectors. Thirdly, run the SVM with different kernels and choose the best kernel function.Then, prefer the best SVM parameters c and g in the on-chip machine using K-fold cross-validation method. Finally, the optimal kernel function and optimal parameters are set as prameters to run the experiment again and the final bleeding image classification accuracy can reach 83% specificity and 94% sensitivity. Compared with other algorithms, the SVM algorithm converges fast, runs stable with strong generalization ability and the method of SVM parameter optimization can be widely applied to other studies of SVM classifier.