Research On Small Intestine Disease Automatic Diagnosis Algorithm For Capsule Endoscopy Images

Wireless Capsule Endoscopy is a new-model endoscopic diagnosis system of gastrointestinal tract, and it has the performance of both micro-camera and video signal transmitter. Since this system has been applied into clinic around the year2000, clinicians can obtain the whole approximate40-60thousands images of the entire small intestine without bringing any pain to patients during per examination. The traditional diagnostic method that only depending on clinicians’naked eyes is time-consuming and labor-intensive. It is necessary to develop a computer-aided system to alleviate the burden of clinicians.The main work of this dissertation is to study out an automatic diagnosis algorithm in order to automatically filter the abnormal images from the giant number of images got from patients. The main idea is to extract color and texture feature of every image, and then using machine learning method to learn and classify the features. The author’s main research works are as follows:Firstly, quasi-invariant which based on the theory of the Dichromatic Reflection Model is a kind of color derivatives obtained by projecting derivative of image on the direction of hue, and it can represent the edge information of object by removing the edge information caused by variance of shadow-shading and high light, but it depends on a physical parameter. Full invariant which can be computed by normalizing quasi-invariant with a signal dependent scalar is more desired for feature extraction. In this dissertation, a full invariant based feature is used as the feature of capsule endoscopy image getting rid of the edge of shadow-shading and high light. At the same time, some color features such as color moment and color histogram and some texture features such as Contourlet transform and local binary pattern are adopted as the feature of capsule endoscopy images in this dissertation. Three classifiers are used to classify the images. Through experiment and comparison, the character of every feature extraction method is analyzed.Secondly, Multi-Instance Learning is introduced to diagnose disease. In this dissertation, an image is divided into several patches. Feature is extracted from every patch. The feature of a patch is an instance, and the set of all the features of the patches from an image is a bag. In machine learning, only the bags are labeled without labeling the instances. After Multi-Instance Learning analysis, both the abnormal images and abnormal patches can be identified. Color moment, color histogram and local binary pattern are used as the feature of image patches. The local features of images are classified by three Multi-Instance Learning algorithms. Though experiment, the performances of above feature extraction algorithm and Multi-Instance Learning algorithm are compared.Thirdly, to cut down the rate of missed diagnosis, a cascade image classification method is used in this dissertation. The normal images identified in the first level may include some abnormal images. In the second level classification, the normal images identified in the first level classification are filtered again to filter out the abnormal images misdiagnosed in first level as far as possible and cut down the rate of misdiagnosis. The experiment result has illustrated that the sensitivity is improved arid the rate of misdiagnosis is declined.