Study On Tumor Detection And Its Preliminary Location Of Gastrointestinal Capsule Endoscope Images

Gastrointestinal cancer is the most common malignancy, which is a serious threat to human health. Its incidence among the top of all malignant tumors, and the patients become more and more younger. How to more diagnose gastrointestinal tumors efficiently becomes a social problem to be solved. The traditional way to detect tumors is invasive, and patients must endure some pain. Gastrointestinal endoscopy capsule is a revolutionary breakthrough in the field of detection of gastrointestinal, which is painless, simple, non-intrusive. However, there still exist many problems in the capsule endoscope now. The first one is the huge number of images produced by capsule endoscopy. Secondly, the position of the image of the digestive tract lesions may appear very few, the doctor's negligence can cause serious consequences of misdiagnosis. For this reason, studying on tumor detection and its localization of gastrointestinal capsule endoscope image will enhance the usefulness of gastrointestinal capsule endoscope.In this paper, the study is based on tumor detection and its initial localization of gastrointestinal capsule endoscope image. The main two parts are as follows:1) Tumor detection of gastrointestinal capsule endoscope image. In this part we take full use of the shape of the tumor to segment capsule endoscope image. And we propose a classifier based on the texture feature of amplitude spectrum to judge the difference between the tumor tissue and the "pseudo-tumor". Usually tumor tissue in the capsule endoscope image is brighter than the surrounding normal tissue, and its shape is similar to an ellipse. Based on this, we enhanced the contrast in the image to makes the tumor more prominent. Then we use the level set method for image segmentation, and use ellipse fitting algorithm to find out areas whose shape like an ellipse. However the results include tumor site, there will be part of the "pseudo-tumor". To solve this problem, we utilize the classifier we proposed to determine whether the tumor exists or not.2) Tumors' initial localization of gastrointestinal capsule endoscope image. In this part we improve the original Weber local features(WLD), make it more robust to noise. And we also propose a two-stage method for locating the pylorus. Through tumors' initial localization we can know which digestive organs(esophagus, stomach, intestines) the tumor locates in. Technical difficulties of this section lies on determine the cut-off point of digestive organs. The locating of cardia is relatively simple. We extract HSI color feature of the images, based on the number of valid child within the image area to locate the position of the cardia. Pylorus' locating algorithm is divided into two phases. During the rough locating phrase, according to the image will be short-term changes in the color when the capsule went into the adjacent digestive organs, we use color characteristics to obtain candidate positions of the pylorus. During the exact locating phrase, we combine K nearest neighbor(KNN) classifier with the improved WLD histogram to select the exact position of the pylorus in the candidate positions by using a sliding window mode. The localization of cardia and pylorus determines the accuracy of tumor's initial localization.