With further development of the computer technology, image processing, pattern recognition and artificial intelligence technology, the whole society is becoming more intelligent. In the medical field, a variety of high-capacity, high-resolution medical imaging equipments are applied. In a medium-sized hospital, the growth rate of CT scan images produced by a variety of medical instruments is more than 10G byte per day. Doctors have to do a lot of diagnosis facing computer screen or films every day. As a result, the efficiency improvement of large-scale usage of inspection equipments is not obvious, and it may lead to unnecessary misdiagnosis and missed diagnosis. The traditional manual way of reading film imags has significantly lagged behind the rapid development of technology and equipment, and with the growing application of greater capacity, higher resolution, more advanced imaging equipment, this contradiction will become increasingly prominent.Medical image is a kind of natural image and has the features of structural continuity, and grayscale consistency in CT images et. al. Theoretically it can be machine identified using image processing technologies and pattern recognition technologies. In long term, to carry out intelligent computer-aided diagnosis based on image processing and analysis will become a major trend. In order to achieve the goal of computer intelligent diagnosis or computer aided diagnosis, the problem first of all must, thus, be faced and solved is automatic detection of lesions. Nowadays, lesion detection based on whole brain scan data especially detection based on texture features is just in the beginning. In this dissertation, aim at the direction of computer intelligent diagnosis, and intensive studies lesion detection methods based on high-resolution 3D cerebral CT images, we put forwards several innovative algorithms or improvement algorithms. Experiments are performed to verify the overall framework and algorithms we proposed. The main research work and contribution of this dissertation are as follows:(1) Proposed the main framework of lesion detection based on three-dimensional cerebral CT data.Currently, lesion detection based on three-dimensional CT volume data is still in the initial stage, there is no unified framework and standards. This paper presents a lesion detection method based on three-dimensional cerebral CT data and feature vectors statistical atlas, and the flow chart and main framework of lesion detection, including data preprocessing, rigid registration and non-rigid registration, atlas creation, feature extraction, and, ultimately, the process of disease detection. We also have done experiments in every stage to verify the methods and the main framework.(2) Proposed a whole brain image segmentation method based on prior knowledge and the continuity of the brain structure, and also an interpolation algorithm in layers.Based on the prior knowledge of cerebral CT images and the continuity of brain structure, we proposed a whole brain image segmentation method which begins from the centure layer and respectively to the base and top of the skull, and it can automatically segment the whole brain images in one time. This dissertation proposed an improved inter-layers interpolation algorithm, which should select a neighborhood window, and then select corresponding point from neighborhood window based on feature vectors. Experimental results show that the proposed method satisfied the interpolation constraints, the image interpolated has clear structure, and retains the structural characteristics and textural properties, and meets the requirement of subsequent data processing.(3) Proposed an improved rigid registration algorithm, to avoid falling into local optimal solution.As Powell optimization algorithm is a local optimization algorithm, it is easy to fall into local optimal solution. In addition, the characteristics of the image itself and the local similarity caused by interpolation are also easy to make the registration process fall into a local optimum. In this dissertation, we proposed an optimization algorithm based on uniform design and Powell. The experiments verified that it can avoid falling into local optimal solution.(4) Proposed the improved Demons algorithm to strengthen the topology preservation of registration.The existing research results can not completely eliminate the non-topology preservation in non-rigid registration. So, we designed an improved algorithm. This is based on the viewpoint that it’s not only necessary to reduce the deformation, but sometimes to increase the deformation for topology preservation. In this algorithm. we maintain the direction of deformation and make two-way optimization of deformation magnitude to make the Jacobian determinant greater than zero. Experiments verified that the improved method can eliminate all deformation points without topology preservation.(5) Proposed a texture feature vector construction method for building statistical mapping and lesion detection.In order to detect lesions with texture features, a texture feature vector with low-dimension and high classification capacity should be constructed. In this dissertation, we proposed several optional texture characteristics based on theoretical analysis, and then do experiments to select the best combination. Experiment verified that the lesion detection result is fine based on the selected texture feature vector. Compared to other high-demension feature vector, it has the features of simple construction, less calculation and good texture classification capacity.We appreciate the support of Nature and Science Foundation of China (Project No.60771007).