A Study On Key Technologies For GPU-Based Medical Data Visualization

The visualization of medical data, which reconstructs a large number of two-dimensional imagesinto a realistic three-dimensional model, plays an important role in the field of clinic diagnosis,surgery navigation and guiding treatments. The fundamental algorithms of visualization are of twotypes: surface rendering and volume rendering, of which volume rendering is more widely used as itallows unbiased visualization of fine details and parallel processing. Based on deep study of the keyfactors that influence the quality and speed of volume rendering, several improvements are made tomeet the special requirements of medical data visualization.Transfer function is one of the most important factors that affect the quality of volume-renderedimages. A well-designed transfer function can highlight the regions of interest. The existing onedimensional transfer function is simple but difficult to extract complex features, whilemulti-dimensional transfer function requires the user to select more parameters, and increases thecomplexity of the interface. In this thesis, data analysis and image-guided transfer functionapproaches are combined. Boundaries of different objects are pinpointed through the analysis of adata structure, which we term the histogram volume. The histogram volume captures therelationships among data values, their first order and second order directional derivatives throughoutthe medical volume data. The transfer function which assigns different optical properties to differentboundaries can distinguish the complex tissues very well. Additionally, a user friendly stroke-basedinterface is proposed to make the multi-dimensional transfer function design more intuitive andconsistent with the users’ specific observation requirements.Raycasting volume rendering algorithm can produce high-quality rendered images. However,due to the heavy computation burden, the real-time reconstruction of large data sets is hard toachieve. By taking advantage of the powerful computational capacity of modern programmable GPU,we convert the traditionally CPU software-based algorithm to the GPU and implement per-pixellighting effects to create more realistically rendered images at interactive frame rates. Furthermore,frame buffer objects are utilized to solve the problem that the basic raycasting algorithm cannotcorrectly render volumetric datasets and polygons together.Finally, a medical data visualization toolkit, which makes full benefit of GPU, is designed andimplemented by extending the open source toolkit VTK. Experimental results suggest that thedesigned toolkit has good performances both in rendering speed and image quality. It is very efficient to render medical data interactively on general PCs.