A Computer-assisted Diagnosis Oriented Research On Image Process And Analysis Of Capsule Endoscopy

Wareless capsule endoscopy, which is a gastrointestinal disease detection tool with benefit of painless, safe and full inspection, has become one of the focuses in clinical research. Capsule endoscopy produces a lot of color images. In order to assist doctors detect the disease, image processing technology is gradually emerging in this field. According to the characteristics of capsule endoscopy images, image enhancement, segmentation and classification are studied in this article.When wareless capsule endoscopy comes through the human gastrointestinal tract, it can not control the location and angle of shooting. The images are greatly influenced by light condition, which doesn't facilitate the medical observation. In this paper, we propose a method based on Retinex theory with total variational model for image enhancement to improve the viewabilty. Based on the analysis of human vision system, Retinex theory deals with the removal of unfavorable illumination effects from images. The algorithms of this theory are widely used in image enhancement for their good performance on color constancy, contrast enhancement and exhibition details in shadow regions. The split Bregman method is employed to calculate this model with extreme efficiency. Beyond fast processing, this model is also good at keeping texture details and color information of the image. We compare the proposed method with two popular methods on the enhancement performance and calculation efficiency in the experiments. The experiment results show that the proposed method is comparable of enhancement performance and nearly 40 times faster than the original method. This method is also applied for bleeding detection in capsule endoscopy images, and a significant result is obtained.By unique filming of capsule endoscopy and special organizational structure of gastrointestinal, the image is very hard for segmentation for blur edges, uneven lighting, and complex background. A new geodesic active contour combined region and edge information for image segmentation is proposed. Motivated by the fact that edge-based models are clever at accurate segmentation while region-based models are less sensitive to noise and have better performance for images with weak edges or without edges, the new model shares the advantages of them by a simple modification of the Geodesic Active Contour (GAC) model. Experimental results demonstrate the efficacy of the proposed model. Also the above algorithm get a better result than other classic active contour style algorithms, it also suffers the shadow of images caused by the bad lighting condition.An other new model is preposed to solve this problem by omitting the region information. This model consists two parts.The first part is an initial region performed as an anchor to identify the position around the target. The second part is the gradient image to identify the positon of the edges. A total variation model is used to organize this two parts to form a convex functional which has a global minimum. This model yields encouraging experimental results on a number of capsule endoscopy images when comparing with many other image segmentation algorithms.The capsule endoscopy images are quite different between different cases, but changed slowly alone the image querry in a case. This means it is very hard to learn the whole structure of centain kind of capsule endoscopy images among different cases, such as small intestine images, but the similarity between neigbered images is useful. This leads a study on semi-supervised learning for capsule endoscopy image classification problems. Graph-based semi-supervised methods define a graph where the nodes are labeled and unlabeled examples in the dataset, and edges reflect the similarity of examples. These methods usually assume label smoothness over the graph. Graph methods are nonparametric, discriminative, and transductive in nature. These methods take high classification accuracy on variant data distributions. But the computation complexity is very high. As the size of dataset growing, the graph will be too large to compute and this limits the extension of its usage. In this paper, we propose a novel method for fast computation based on local clustering, which is very efficiency for reduction the size of graph and at the same time the accuracy can be maintained. The local clustering method is low in computation complexity and the data structure can be preserved by a new designed distance function. Experimental results show that this approach preserves the accuracy of purely graph-based methods and at the same time significantly reduces computational cost.