Research On Kidney Segmentation And Renal Lesion Detection In CT Images

Based on the evolving progress of medical imaging technology and computer vision, medical image processing and analysis play increasingly prominent roles in modern health care. It has become effective techniques in clinical diagnosis, treatment planning and surgical navigation, which is valuable for the development of computer-aided detection and diagnosis system in medical application. As one branch of medical image processing and analysis, medical image segmentation is always the key and difficult issue in the field. Fast and effective segmentation methods can provide doctors with reliability for patient treatment and diagnosis, thus improving the efficiency and accuracy of health care.Considering the imaging characteristics and clinical requirements of medical images, this thesis is to research deeply on related science technology about medical image segmentation and lesion detection for kidney CT images that are regarded as study objects, by summarizing the research results achieved both at home and abroad. Our main contributions can be summarized in the following aspects:1. Because of the imperfection in dealing with intensity inhomogeneity by traditional C-V model, an improved C-V model is proposed to handle this problem for kidney segmentation, according to the property of kidney tissues in medical images. This algorithm integrates global and local statistical imformations into C-V model, which makes it more reliable for image segmentation and effectively eliminates the influence of intensity inhomogeneity.2. On basis of kidney characteristics in CT image sequences, we propose a new graph-cuts-based active contours model with an adaptive width of narrow band for kidney localization and extraction. The adaptive search range for evolution of the energy function is determined by combining contextual continuity with the object size. Then, based on segmentation results of the middle two slices in the CT image sequence, the proposed model is applied sequentially on the segmentation of the remaining slices. The energy function combines geodesic active model (GAC model) and C-V model, which takes into account boundary and regional informations. In addition, the corresponding weights of t-link are converted into weights of n-link for reducing graph edges. Due to the proper area for evolution, the opertational range of the active model with narrow band is limited to ensure computational efficiency.3. In order to make up for the imperfection of interaction segmentation in entire operation time, a fully automatic algorithm for kidney segmentation is proposed on the basis of the above model. As the reference image, the middle slice is automatically segmented by renal cortical property and C-V model for initial segmentation of the whole CT sequence. According to the empirical relationship of the shape difference in adjacent slices and corresponding layer thickness, the proper width of narrow band is calculated to benefit the rapid evolution of the energy function.4. The framework of renal lesion detection is constructed on account of support vector machine, including sample selection, feature extraction and establishment of the classification model for detection. Based on the CT performance of kidney lesions, the method uses statistical search to obtain candidate sample regions with the circular window template and then extract the intensity and texture characteristics, which can help doctors detect potential kidney lesions.