Segmentation Of The Cerebral CT Images Based On Registration

While medical images of high resolution and high capacity provid a powerful tool for clinical diagnosis, it also increases the workload of radlologist. In this context, Computer Aided Diagnosis has become the major challenge and main direction in the field of medical imaging processing and analysis.With the goal of pathology automatic detection on the cerebral Computed Tomography, we have made a in-depth study on the registration and segmentation of the CT cerebral images. We made some success, which do preparations for the target of realizing pathology automatic detection.Non-rigid registration is not only the key technology for the realization of pathology automatic detection, but also the necessary step of constructing the deformation atlas of the normal brain anatomy. Demons algorithm is a accurate non-rigid registration algorithm, but it is derived by assuming small deformations.So one of the limitations of the original Demons is that it can not produce topology preserving maps for the large deformations. Aiming to solve this problem, an improved Demons algorithm was proposed in this paper. First, the equation of force in the original Demons was regarded as the result of minimizing the energy function. Then, Demons algorithm was improved by adding a regularization term into the function. The symmetric Kullback-Leibler (sKL) distance in information theory was used as the regularization term. The experiment results with high resolution CT cerebral images demonstrated that the improved algorithm could not only handle large deformations, but also obtain accurate registration results using smooth deformation fields.Based on the improved registration algorithm, a registration-based automatic segmentation algorithm on the CT cerebral images was proposed in this paper, which imitates the process of manual segmentation. The goal of the algorithm is to divide the brain into three categories: gray matter(GM), white matter(WM) and cerebrospinal fluid(CSF). This method requires a priori map, which is obtained by manually segmenting the reference image in this paper, to guide the segmentation. Before manually correcting the image in use of anatomical knowledge, the reference image was pre-segmented by using median filter and fuzzy C-means clustering. The process of the segmentation algorithm is as follow. First we use the priori map to segment the registration result, which is obtained by matching the floating image to the reference image. And then the segmentation result of floating image is gotten by inverse mapping. In the experiment, we improved the final results by cluster segmentation and mathematical morphology. The experimental results show that the algorithm although has some limitations, it is feasible.