School of Computing
50 S Central Campus Dr, RM 3190
Salt Lake City, UT 84112
Real-Time Volume Rendering of 3D Medical Scan Data
M.S. Thesis - Cornell University, 2010
Abstract:
The immense progress of imaging technologies has radically changed the practice
of medicine both in terms of diagnosis and intravascular surgery. Using
technologies such as magnetic resonance imaging (MRI), and computed axial
tomography (CAT) scans, doctors are now able to "see" internal organs and
structures in high-resolution detail. Today using expensive specialized hardware,
one can generate three-dimensional visualizations providing accurate interpretations
and revolutionizing the medical field.
This thesis presents a substantially different method to visualize volume
datasets by treating them as a scattering volume and rendering the images on a
small cluster of parallel computers. With sufficient computing power, the data
can be explored interactively without any loss of information.
We utilize a basic raycasting algorithm with several acceleration techniques,
such as global empty space skipping, early ray termination, a global gradient
cache and increased data access coherency. By selecting efficient data subdivisions,
we eliminate the memory and bus-bandwidth latencies and maximize the
computing power of each core. The cache coherence of the data access due to
the bricking scheme produced almost real-time rendering speeds that are independent
of the viewing direction. We tested these algorithms on three different
datasets at varying output image resolutions.
In the near future, with increased computing power and sufficient bandwidth,
it will be possible to use a cluster of machines to render time-dependent
datasets in real time and to deliver these images directly into an operating room.
Files:
document (pdf, 20.0 MB)
BibTex (bib, 250 B)
presentation (pdf, 3.71 MB)
Media: Videos were created to illustrate the speed of the entire system.
We tested three datasets (930 and 770 slices at the resolution of 5122). Output images
at the resolution of 10242 were generated at the rate above 10 Hz, with the actual refresh
rate shown in the top right corner of the videos. We used 256 cores in parallel, each running at 2.6 GHz, to compute each frame.
The left video is at half resolution (5122) (H.264, mp4, 14.5 MB) and the one on the right is at full resolution (10242) (H.264, mp4, 39.0 MB).
Images: Pre-Operation Dataset
Post-Operation Dataset
CT 14 Dataset
Acknowledgements:
This work would not have been possible without the support from
Don Greenberg,
Alex Vladimirsky,
and everyone in the Program of Computer Graphics.
This work was supported by the
National Science Foundation ITR / AP: CCF-0205438, the Department of Architecture, and the
Department of Computer Science.