The Implementation Of GPU-based Multi-function Visualization Of Medical Images

Medical image visualization is an important application of visualization technology in the field of medical image processing, and it is also the research hotspots content of the current medical image processing. The three-dimensional images reconstructed by the medical image visualization can make up for the lack of the three-dimensional image which is produced by the medical imaging equipment by displaying the complex internal structure of the human tissue visually. Volume rendering as the key medical technology of image visualization has been the rapid development in recent years. Compared with the standard three-dimensional surface rendering, volume rendering has the biggest advantage that it can provide a translucent rendering which shows rich spatial information among different structures. As the classic algorithm of volume rendering, ray casting algorithm can produce more realistic, high-quality images because of considering all of the voxel data contribution to the image and using as much as possible of the original information during the process of rendering. At the same time, the amount of data generated by the modern medical imaging equipment is very large, and in order to produce undistorted images each ray requires a large number of sampling points in ray casting algorithm, it results in particularly large amount of computation and real-time displaying of the effect can’t reach, so some chooses to sacrifice the image quality for a good interactive performance, or abandon the interaction for a high-quality of the reconstructed image. With the development of science and technology, the image processor GPU (Graphic Processing Unit) computing speed and programmability has been significantly improved, under the premise of ensuring the quality of the image, the researchers have solved the problem of the interaction by improving and designing a lot of algorithms according to the special structure of the graphics hardware architecture. Based on Learning GPU programming, we have implemented the real-time display of the ray-casting algorithm on an ordinary PC and done some work of multifunction display of medical images.1.Accelerated Ray casting algorithm based on CUDAIn the paper we use the GPU programming language CUDA (Compute Unified Device Architecture) to give full play to the powerful parallel computing ability of the GPU.We put the tri-linear interpolation calculation and integration process which is time consuming in ray casting algorithm on the GPU, thus greatly saves the time. Firstly, use the DCMTK which is the development kit of digital images and communication standard DICOM (Digital Imaging and Communications in Medicine) to read the images having DICOM format into memory. Then map the transfer function and the voxel data in the memory for texture and load them into the Graphics memory by invoking the CUDA memory operating function. Then, according to the CUDA programming model we use the grid to simulate the entire viewing plane. A thread of the grid which can be determined according to the thread index value corresponds a pixel in the view plane. The light starts on the viewpoint position, and considers the difference between the pixel coordinates on the view plane and the viewpiont coordinates. In order to reduce the amount of unnecessary calculation and improve computational efficiency, we use the method of seeking the intersection of the light and the box which is bounding the voxel data for the part outside of the box is useless to the rendering image. For the intersection, the image is composited from fiont to back to accelerate the speed. The final rendering image is displayed directly by the interface of OpenGL (Open Graphics Library) and CUDA. Experimental results show that the method of rendering speed can meet the real-time needs of the medical image visualization, and has good prospects for clinical application.2. The implementation of GPU-based multi-function visualization of medical images3D visualization systems using a variety of human-computer interaction operating on the image makes the doctors observe and analysis from the multi-angle and multi-level to facilitate the doctor more intuitive to see the internal complexity of the structure of the human tissue on the screen, so it can help doctors diagnose more accurately or design a more accurate surgical planning.On the base of the algorithm that Accelerated Ray casting algorithm based on CUDA, the paper uses Visual Studio2008as the development platform and designs the interface by MFC (Microsoft Foundation Classes). Firstly, read the DICOM image into the memory by DCMTK, then save the volume rendering data and color, opacity values and other optical properties to the texture memory of the GPU. In order to accelerate the speed, access the data through the cache of the texture memory and hardware accelerate the interpolation calculation of resample points. After that use OpenGL and CUDA interoperability to implement the export of the volume rendering.In order to better determine the location of the structure of interest, the user can select the part of interest by drawing a line using MFC in the graphics device interface GDI (Graphics Device Interface) in the three directions of the two-dimensional image of the cross-sectional, sagittal plane, coronal plane. Firstly, get the position of the cursor when the left mouse button clicking events happens by using the cursor position obtained function in the MFC. Then determine if its location is in three sections. If it is true, then save the data as the coordinates of the line and draw the line using the GDI function. Then convert the coordinates to the three-dimensional volume data space. For the sampling points on the line, give it a brighter color and larger opacity, so the difference between the point on the line and the surrounding tissue can be displayed obviously.In order to overcome the defect of the performance on the spatial position of the two-dimensional tomographic image, embed the two-dimensional tomographic images into the three-dimensional image. So the doctors don’t need to imagine the structure of the human tissue location when they observe the two-dimensional tomographic image. First by a left mouse click to draw a line in the three directions of the tomographic image, then through coordinate conversion to the three-dimensional data field space and a point outside the line is defined by the right mouse to define a plane. During the volume rendering, determine the distance between the resampling points to the plane defined, if the distance is in the range give the point the corresponding gray scale values and opacity values. Then proceed with the cumulative calculation of the light projection of the color and opacity values. The final image is outputed by OpenGL and CUD A interoperability and then achieves the two-dimensional tomographic image is displayed in the three-dimensional image.The organizational structures in the three-dimensional image are overall displayed, so that the interested parts within the organization are blocked. So the cutting of the three-dimensional image is proposed that can be seen the internal component of the organizational structure by cutting the area obscured. The blocked region selection is achieved by painting a rectangle in the three-dimensional images. First using the OpenGL paintings rectangle function draw a rectangle in three-dimensional images, then transform rectangle coordinates into the space of three-dimensional data field. In order to achieve the purpose of the resection block part, for the re-sampling points in this region do not participate in the process of ray casting. In order to adjust the size of the occluded region, two sliders that corresponding weight function in both directions are added.When it is displaying the three-dimensional image, the relationship between the organizational structures and blocked ones surrounding is prefer being seen to being cutted. The selection of a particular area is also achieved through the painted rectangle drawn in the three-dimensional image. Choose the layered peeling algorithm of volume rendering for the point in the selection part to achieve its special display.3. Layered peeling algorithm of volume rendering based on Laplacian of GaussianNow, one of the most common problems in medical image volume rendering is occlusion problem between the different tissues in the human body. Generally it is solved by adjusting the transfer function. Better results can be obtained in the case that two tissues differ greatly for example muscles and bones. When the two organizations have little difference, the organization inside can not be seen very well, for example the brain and the surrounding soft tissue, unless use multi-dimensional transfer function and adjust the complex parameter. Currently another approach to solve organizational occlusion is pre-segmentation of the voxel data which is hardly to classify the data with reason or process automatically. Thus, how to peel the interest area with simple operation in real-time effect is important. Using the principle of edge detection to determine the layered point between the different organizations, put forward a method that using Laplacian of Gaussian to determine the layered point of different organizations by judging the zero-crossing of the image second derivative because of the shortcoming that it is difficult to choose the appropriate threshold of the gradient in the layered peeling algorithm of volume rendering based on the gradient. Gaussian Laplacian is composited by a Gaussian filter function and Laplacian operator. The Laplacian is a second derivative operator which does not need to set precise threshold in the extraction of the edge point like the first derivative operator and decides whether it is the edge point through checking its zero-crossing. During the ray casting process, for each sampling point, to decide whether it belongs to the edge point, and if they are the edge points, then this point is hierarchical point; if not, then one step forward in the light direction, determines a next sampling point whether it belongs to the edge of the point to stop until you find the edge points. Since some parts of the edge are not constituted by a single layer of points, in order to achieve a better layered effect, consider the edge points of the same edge as the same stratified layer. At the same time, in order to achieve real-time effects, the graphics processor is used to accelerate the algorithm. The experimental results show that the algorithm can do the layered peeling of volume rendering automatically, clearly and in real-time.In the paper we use the GPU programming language CUD A to give full play to the powerful parallel computing ability of the GPU and implement the ray casting algorithm on an ordinary PC in the real-time with good image quality; We proposed a three-dimensional interactive multi-functional visualization of medical images based on accelerated ray casting algorithm with CUDA which can achieve the function of orientation, any cross-section, cutting and special display in a three-dimensional image. By interacting with the system, it provides a better user experience and diagnostic services for physicians. The experimental results show that on a typical desktop PC, the system which has better real-time and interactive can give good service to the doctor and has great application value; To solve the problem of inter-organizational occlusion, the paper proposes a layered peeling algorithm of volume rendering based on Laplacian of Gaussian with the principle of edge detection. The experimental results show that the algorithm can provide a clear and real-time layered display of volume rendering automatically.