Digital Brain-Research On Methods Of Computational Anatomy And Applications On GPU Technology

With the development of brain imaging technologies, brain imaging is turned to be very important to the study of neuroscience and neurosurgery. The rapid collection of brain image data stimulates the study of methematics methods and computational algorithms to compare, pool and average brain data across whole populations, so as to that a new discipline computional anatomy is emerging. In the brain data domain computional neuroanatomy focuses the study mainly on the brain atlas modeling, deformable model and correlation between brain structures and functions mapping. In this paper, several key problems within computational neuroanatomy are introduced and conducted to the research on the detection and extraction of brain structures and anatomical landmarks, brain image registration technology based on elastic model, and to the initial research on the cortex segmentation based on K-means clustering. At last we show the study of the applications of programmable graphics processing unit(GPU) in the visualization of digital brain and the accelerated medical image processing.In the area of brain image segmentation, firstly for the non-brain and brain segmentation two methods are studied in which one is based on edge segmentation and the other is based on watershed method, both get good results but the difference in interaction. Then much effort is payed attention to the brain tissue segmentation in which a fuzzy clustering method with a prior probabilities atlas is studied and the intensity non-uniform is investigated. At last, the anatomical landmark is introduced with its'model and methemathics method for curvature on iso-surface and the experiments of semi-auto detection of landmarks show good performance.For the deformable registration of brain images, an elastic formulation for registration is introduced and the finite-element model is used for the computation. The adaptive FEM mesh is generated according to the gradient field of brain atlas, so the mesh can also represent the features of the atlas. This makes the registration more accurate within the same condition. The landmarks are detected to determine the rigid transformations between corresponding images before FEM is applied and also used for the restrictions of FEM deformation.The cortex volume of brain structure is deeply folded by sulcus on the gyrus which are the very complex in it's morphology but import to the fMRI data and brain function analysis. To segment out the cortex into parcels is very helpful for the regional data analysis in fMRI and the study of the automatic labeling for brain volume. A geodesic based k-means clustering method is applied to the cortex volume and methods of examination for the new centroid are compared. The clusering method produces a good segmentation on the volume data with a fast convergency.For the application of graphics processing unit(GPU) in medical image area, it has been seen the proved value of GPU as a computational resource for both general-purpose computations and tasks that are related to medical imaging. Within a framework of GPU programming developed for visualization and image processing, some methods for high quality volume rendering are compensated and a large-scale volume rendering based on vector quantization are studied for real-time decoding and rendering on GPU. At last, the implementations on GPU for the level set solver for brain image segmentation and skeleton extraction are studied and tested. In general one can say that a GPU-based solution is likely to accelerate computations when they can be done on independent elements, i.e. the computations can be mapped onto a streaming model, because this makes it possible to fully utilize the parallellism of the GPUs.