Mri Brain Tissue Segmentation And Gpu-based Acceleration

With the development of medical imaging technology, high-resolution medical imaging equipments such as magnetic resonance imaging device are more commonly used in clinical diagnosis. It provides convenience for doctors'accurate diagnosis, however, reading the increasing clinical images has also become a heavy burden for doctors. So research the computer-aided-diagnosis system based on medical image is an effective way to solve this problem.Brain lesions like cerebral hemorrhage and brain tumors are common diseases in clinical. So how to make a construction of computer aided diagnosis system to realize the brain lesions automatic detection with the help of medical imaging technology is important to clinical diagnosis and it is also a challenging research job. One makeble idea is creating normal people's brain atlas based on image segmentation and registration technology, and then establish the reasoning system based on brain atlas.Tissue segmentation is one of the basic technologies in medical image processing, and the quality of the results directly affects following process, this is why image segmentation has always been one of the important parts of research. Because the diversity and complexity of given initial conditions, image content and object content, fully automatic segmentation methods can not give satisfactory results in many cases. By contrast, because of the involvement of the operator, the semi-automatic segmentation methods obtain more precise initial parameters, therefore produce more accurate results. Random walk algorithm is one of the semi-automatic segmentation methods, it has a strong ability to identify the weak boundary, and easily expand to 3D image processing.This paper analysis the theoretical properties of random walk algorithm in-depth, and introduce the practical behaviors of algorithm. The weighting function is one of the most important factors that impact the performance of random walk algorithm, so improving the mapping ability of the weighting function will be able to achieve a better performance. In this paper, we introduced the concept of local image dispersion into random walk, to construct a new weighting function that not only reflecting the changing information of adjacent-pixel's gray value, but also the discrete information of local image, it improved the ability of algorithm to identify homogeneous pixels and edges. This paper has also given two methods to compute the local image dispersion, the image local entropy based method and the LBP based method. We also used Fisher discriminate function to calculate the optimal classification threshold. Target classification threshold in the original algorithm is 1/K, K is the number of label mark. This calculation method is simple, however it didn't consider the differences of result probability distribution between the target and the background, so it ofen occurs classification errors. In this paper, we used the optimal threshold to reduce caculation caused by classification error.Medical image processing algorithms are dealing with large volume of data and the processing time is huge, which make a lot of algorithms can not be applied to actual clinical diagnosis. In recent years the rapid development of GPU technology makes it has became a powerful computing platform, and particularly suitable for compute-intensive applications, such as medical image processing. In this paper, we used GPU to accelerate the improved random walk algorithm, and analyzed the performance of GPU acceleration.