Research And Implementation Of GPU-Based Kernel Regression Reconstruction For Freehand 3D Ultrasound

Image-guided technology in actual medical work has been widely used, and ultraso und technology compared to other imaging technology has various advantages of non- invasive, non- ionizing, portable, low cost, so 2D ultrasound technology is currently used by most medical imaging technology. However, the 2D graphics can not provide complete data of organs and tissues. 3D ultrasound technology would become very meaningful. The Chinese Academy of Sciences has done many studies on the reconstruction of 3D ultrasound, they found the Kernel Regression Reconstruction for Freehand 3D Ultrasound algorithm based on nonparametric estimation method has the advantages of high quality of the reconstructed image. Compared with other 3D ultrasonic reconstruction algorithms, the kernel regression algorithm can effectively suppress speckle noise while preser ving the image boundary, but because of its high time complexity, it also has a serious operation time. In order to make the algorithm more practical application, it is necessary to speed up the algorithm, which is the main purpose.After many other attempts, the answer is that the algorithm is suitable for parallel processing. This thesis is based on the GPU programming technique. And the core of this paper is to obtain a CUDA based free kernel regression method, which is based on GP U. The purpose is to improve the algorithm. In order to verify the success of the algorithm, the paper designs a set of ultrasonic data to perform 3D reconstruction operation, which is used to compare the time difference between the CPU version of the kernel regression algorithm and the GPU version of the kernel regression algorithm. At last, in order to further understand the effect of parameters selection on the actual reconstruction, this paper uses the root mean square error method to evaluate the effect of different size of the functional bandwidth on the reconstruction quality, and then obtain the different parameters of different parameters.The important success of this paper has two parts: one is to achieve GPU accelerated processing of kernel regression algorithm, and the algorithm speed up a lot; the other is through a series of experiments can know how to choose the kernel regression algorithm parameters to achieve the best effect. In the actual implementation of the algorithm, the CUDA programming model itself has a local memory size limit for a single GPU thread, so it can't get more accurate image. Form this reason, the further optimize is needed to achieve higher accuracy.