The Research Of Fast Brain Extraction Based On CUDA

With the rapid development of medical imaging technology, Magnetic resonance imaging technology, avoiding radiation on human cells and wounds on human-body,can clearly distinguish soft tissue characteristics with multi-dimensional image output,.Brain extraction from MR images of the brain become a very important pre-procedure in brain function and clinical application of research. In practical applications, in order to detect small differences between patients and normal brain, we use brain extraction to compare details between a large number of patient brain images and normal brain images. Thus, brain extraction requires high extraction accuracy, speed, and stability.Manually brain extraction methods can achieve high accuracy, but are time consuming. The operator needs to be very clear on the anatomy of the brain, and can be influenced by subjectivity. The hybrid method for automatic brain extraction can achieve higher accuracy and stability, which’s processing time, is faster than the manual,however the speed is still difficult to meet the clinic needs. To do this, we propose a fast CUDA-based brain extraction method, using a hybrid combination of the three parallel processing algorithms(improved BET, level set, and region growing) for automatic extraction.In this paper, a multi-core GPU accelerating ways is used to quickly extract brain tissue based on CUDA. First, a parallel BET algorithm is used to obtain a rough brain boundary by way of allocating different threads to concurrently process different vertex on brain boundary. In each iteration, one thread calculates the new position of one surface vertex; one block calculates the new position of all vertexes with one slice. Thus,after N times of iterations, the new positions of all the vertexes in K slices are updated at the same time. In order to improve the accuracy of the algorithms, we use the resultant edge from the current slice to initialize the contour in adjacent slices by shrinking the result contour of the current slice. When processing 110 MR images of the brain slices, experimental results show that the parallel BET method for brain extraction is global optimum in the accuracy and stability, with running time less than 1 second.Discrete vertexes obtained by BET are connected into closed contour to get the initial outline of the brain, and then the initial mask of each slice brain tissue can be obtained from the outline by region growing method. Finally, we use level set andregion growing combination methods to extract accurate brain tissue. Experimental results show that the level set method for parallel processing is local optimum in accuracy and stability, when processing an MR image of the brain slices, the running time less than 1 second for one slice, this hybrid method based on CUDA parallel processing achieves a rapid extraction of the brain tissue.