University of Utah Interactive Processing of Level Sets 
Aaron Lefohn and Ross Whitaker
Scientific Computing and Imaging Institute

Abstract

    The goal of this project is to accelerate the level set deformable surface methods to interactive rates by mapping both the computation and visualization to programmable graphics hardware (GPU).
Results

1. A 3D GPU-Based Level-Set Solver with Curvature Flow:



We have implemented a 3D level set solver running on the ATI Radeon 8500 GPU. In addition to running in 3D, the solver computes the second-order, mean curvature speed term in spite of the 8-bit memory limitations. The GPU-baed solver runs at 1x-2x faster than a higly-optimized sparse-grid CPU-based solver. We demonstrate the solver segmenting the cerebral cortex surface from a 256 x 256 x 175 MRI volume.
Technical Report:
    Aaron E. Lefohn,Ross T. Whitaker, "A GPU-Based, Three-Dimensional Level Set Solver with Curvature Flow," University of Utah School of Computing Technical Report, UUCS-02-017, (2002)

2. A Streaming Narrow-Band Algorithm: Interactive Computation and Visualization of Level Sets:



These IEEE Visualization 2003 and IEEE Transactions on Visualization and Computer Graphics (TVCG) papers describe a GPU-based, interactive level-set computation and visualization system. The new streaming, narrow-band level-set solver runs at 10-15 times faster than a highly optimized CPU-based implementation. We demonstrate the system configured as an interactive volume segmentation application using the ATI Radeon 9800 Pro GPU.
Web site for papers
    Aaron E. Lefohn,Joe M. Kniss, Charles D. Hansen, and Ross T. Whitaker

3. Brain Tumor Segmentation User Study:



This MICCAI 2003 paper describes an evaluation user study performed with the interactive segmentation application built with the GPU-based, streaming narrow-band level-set solver. Our users segmented nine of the brain tumors from the MRI data set available from Harvard Brigham and Women's Hospital. Our results show that compared to expert hand contouring, our technique is comparably accurate, more precise, and much faster (6 minutes per segmentation versus hours).
Web site for paper
    Aaron E. Lefohn,Joshua E. Cates, and Ross T. Whitaker