Abstract:
HTML
Purpose: To leverage the programmability of modern graphics cards to segment volume data interactively with human guidance in the form of positive and negative "seeds" for hinting at desired and undesired structures, respectively.
Methods and Materials: Our segmentation problem is mainly posed as one of growing structures that the user initializes with seed data that indicate to the algorithm the desired anatomical structure to segment. We implemented the entire segmentation algorithm, including the volume rendering visualization of the segmentation, on the ATI Radeon 9800 Pro video card using the ARB fragment program extensions to OpenGL and the card's support for render-to-3D-texture. The algorithm is a seeded region growing algorithm with a forward Euler discretization of the Perona and Malik nonlinear diffusion equation as a merging metric for the seeds. We also implemented a computation masking program on the graphics hardware that avoids computation for voxels unlikely to change state in the current iteration. During iteration, the algorithm continuously produces a shaded volume rendering of the volume and segmentation data, thereby allowing the user to observe and/or interact with the evolving segmentation by painting additional positive or negative seeds.
Results: At maximum throughput our algorithm processes 60 MVoxels/s and sustains 3.4 GFlops, resulting in a total processing time of 33 ms per 2 MVoxel volume. The computational masking program boosts efficiency to achieve 90 MVoxels/s, which requires only 22 ms per 2 MVoxel volume. For comparison, running only the diffusion without performing the actual seeded region growing on a P4 2.4 GHz CPU with a 533 MHz front side bus and 1066 RDRAM sustains only 1 MVoxels/s, i.e., is almost 2 orders of magnitude slower. When simultaneously displaying the progress of the segmentation using a 3D texture hardware-based volume renderer, Blinn-Phong shading, and calculating gradients on the fly for lighting, we were able to render the 2 MVoxel volumes between 40-100 ms per frame. For comparison, the throughput measured from the diffusion-only algorithm run on the CPU would limit the visualization of a segmentation iteration to 2 seconds per frame, which is at least one order of magnitude longer.
Conclusion: Modern graphics cards can be programmed to achieve interactive segmentation with simultaneous visualization and user monitoring and guidance. This approach obviates the need for completely general and automatic segmentation, a goal that has been and continues to be elusive.
Sherbondy MS, A,
Interactively Guided Volumetric Segmentation Using Programmable Graphics Hardware. Radiological Society of North America 2003 Scientific Assembly and Annual Meeting, November 30 - December 5, 2003 ,Chicago IL.
http://archive.rsna.org/2003/3107579.html
