Unsupervised Segmentation And Local Feature Matching Based Lymph Nodes Detection For Stomach CT Images

Gastric cancer is one of the most common tumors, and it has the highest incidence among all of the malignant tumors in our country. Stomach CT image is a frequent-used measure for the detection of gastric cancer. Clinically, the diagnosis of gastric depends largely on the lymph nodes’metastasis, then, detection of the lymph nodes in the region of interesting correctly is important. However, the traditional detection of lymph nodes is implemented by doctors’manual comparison, which is unreasonably time-consuming, and largely depend on the doctors’ professionalism and cannot meet the demand of medical research development. Therefore, with the rapid development of computer technology, it is indispensable to develop an effective method to realize the automatic detection of lymph nodes.In this thesis, we divided lymph nodes detection into three parts:the preprocess, the detection of suspicious lymph nodes and the recognition of lymph nodes. This thesis gave the following research:Given two unsupervised segmentation based suspicious lymph nodes detection methods.-(a)For images not including the adherent lymph nodes, considering the lymph nodes generally have much smaller than the area of other organs, window of a certain size is used to detect the suspicious lymph node, so windows composed by pixels around the lymph nodes have relatively poor regional coherence, and windows near the other organs do not. Firstly, a sliding window is used to mark the approximate location of consistent region and region of interest (Rol); Then the OTSU multi-threshold segmentation algorithm automatically select seed points in the consistent region for region growth, thus obtaining the integrated consistent regions, and the remain part in the preprocessed image is Rol, which including fat, lymph nodes and some noise. Finally we use marked watershed and feature extraction to get suspicious lymph nodes;(b)For images containing adhesions lymph nodes, a method based on ellipse fitting is proposed to detect the suspicious lymph nodes, lymph nodes has smaller area and faster-changing edge direction, so we combined circular sliding window marking with OTSU binary segmentation to obtain edge points of interest, and used ellipse fitting close contours.Proposed a local feature matching based multi-target tracking algorithm, and designed a series of performance evaluation indexes. Since the position,size and shape of suspicious lymph nodes changes slowly in the stomach CT image sequences,the dissertation restricts the search area in a local window, then by definition and calculation of cost function between the tracked goals and the to be matched goals in the local window, we find the new matching goals,so we can avoid the bug tracking as much as possible. Experiments are conducted on artificial marked images from10cases, results shows that the novel lymph nodes detecting methods are effective.