A method for fast navigation inside a high-resolution 3D radiographic image has been presented. The main feature of the scheme is the reduction in computational cost during navigation; this arises from: (1) the efficient viewing geometry used; (2) the computation of a data structure based on the ADRT that enables the rapid generation of multiple views at a viewing site, and (3) the incremental updating of views from one viewing site to the next. The results demonstrate the practicality of the navigation scheme in performing dynamic and interactive exploration of large 3D radiological images.
This paper only describes the rendering method based on volumetric compositing. But the method can easily be adapted to other simpler rendering schemes, which are typically more than sufficient in most applications. This would result in still further computational savings. (Simpler rendering schemes should be used, for example, if the image is presegmented into different regions.)
Surface rendering, which we do not address, has been popular in 3D rendering. Surface rendering has been made very fast via special hardware [15] and algorithmic considerations. It has received some attention in 3D medical imaging, but more recent work has favored the volumetric approach (e.g., [33,36]); this is partly true, because (1) the volumetric approach does not require presegmentation of data and (2) it is well-suited to the complex ``unstructured'' form of anatomical features. Because of our intended application -- exploration of complex unknown local anatomical data, we do not see surface rendering as an appropriate rendering approach. Geiger and Kikinis, however, have done some preliminary work on an endoscopic simulation/rendering system that uses surface-based techniques, but requires presegmentation [37]. But this system does not work at interactive speeds and it uses no special method for dynamic surface rendering.
Although we have considered only parallel projections, it may be possible to extend the ADRT sampling technique to obtain a data structure for fast perspective projections as well. In this case, the front face of the viewing pyramid is collapsed to a point, and the back face remains a plane perpendicular to one of the coordinate axes. All rays would still be sampled as defined by the ADRT. Rotations are simple in this geometry, requiring only that a plane of rays be added to one side of the pyramid and removed from the opposite side (similar to sideways and up/down translation in parallel projection). However, in/out movement becomes more complicated and would require some modifications to the data structure if recomputation from scratch is to be avoided on each incremental in/out step.
We have also begun applying this concept to the exploration of the 3D coronary arteries [8]. Our long-term goals are to have a means to visualize and make measurements on complex branching structures in 3D pulmonary images and 3D coronary angiograms [9]. The method described here is being integrated into a GUI system for endoscopic visualization of 3D radiological images. This system will not only permit measurements, but also have movie making capability and more convenient means for invoking navigation [38].