Interactive, GPU-Based Level Sets for 3D Segmentation Aaron Lefohn Joshua Cates Ross Whitaker University of Utah Problem Statement Goal • Interactive and general volume segmentation tool using deformable level-set surfaces Challenges • Nonlinear PDE on volume • Free parameters Solution • Accelerate level sets with graphics processor • Unify computation and visualization University of Utah University of Utah Level-Set Segmentation Surface velocity attracts level set to desired feature % Smoothing Data-Based Speed Curvature Speed Segmentation Parameters 1) Intensity value of interest (center) 2) Width of intensity interval (variance) 3) Percentage of data vs. smoothing University of Utah Data speed term Attract level set to range of voxel intensities Width (Variance) Center (Mean) D(I) D(I)= 0 I (Intensity) University of Utah Curvature speed term Enforce surface smoothness • Prevent segmentation “leaks” • Smooth noisy solution Seed Surface No Curvature With Curvature University of Utah Why GPU-Based Level-Set Solver? Inexpensive, fast, SIMD co-processor • Cheap (~$400) • Over 10x more computational power than CPU • Fast access to texture memory (2D/3D) Example GPUs • ATI Radeon 9x00 Series • NVIDIA GeForceFX Series University of Utah General Computation on GPUs Streaming architecture Store data in textures ForEach loop over data elements • Fragment program is computational kernel Texture Data CPU Vertex & Texture Coordinates Vertex Processor Rasterizer Fragment Processor Frame/Pixel Buffer(s) University of Utah GPU-Based Level-Set Solver Streaming Narrow-Band Method on GPU • Multi-dimensional virtual memory • Optimize for GPU computation – 2D, minimal memory, data-parallel Virtual Memory Space Physical Memory Space Unused Pages Inside Outside Active Pages University of Utah Evaluation User Study Goal • Can a user quickly find parameter settings to create an accurate, precise 3D segmentation? – Relative to hand contouring Methodology • Six users and nine data sets – Harvard Brigham and Women’s Hospital Brain Tumor Database – 256 x 256 x 124 MRI • No pre-processing of data & no hidden parameters • Ground truth – Expert hand contouring – STAPLE method (Warfield et al. MICCAI 2002) University of Utah Evaluation Results Efficiency • 6 ± 3 minutes per segmentation (vs multiple hours) • Solver idle 90% - 95% of time Precision • Intersubject similarity significantly better Accuracy • Within error bounds of expert hand segmentations • Bias towards smaller segmentations • Compares well with other semi-automatic techniques – Kaus et al. 2001 University of Utah 3D User Interface Demo QuickTime™ and a Video decompressor are needed to see this picture. University of Utah Conclusions 1. GPU power interactive level-set computation • Streaming narrow-band algorithm • Dynamic, sparse computation model for GPUs 2. Interactive level-sets powerful segmentation tool • • • • Intuitive, graphical parameter setting Quantitatively comparable to other methods Much faster than hand segmentations No pre-processing of data & no hidden parameters Future work • Other segmentation classifiers • User interface enhancements More information on GPU level-set solver • See IEEE TVCG paper, “A Streaming Narrow-Band Algorithm” • Google “Lefohn streaming narrow” University of Utah Acknowledgements Joe Kniss Gordon Kindlmann Milan Ikits SCI faculty, students, and staff John Owens at UCDavis ATI Technologies, Inc • Evan Hart, Mark Segal, Arcot Preetham, Jeff Royle, and Jason Mitchell Brigham and Women’s Hospital Tumor Data • Simon Warfield, Michael Kaus, Ron Kikinis, Peter Black, and Ferenc Jolesz Funding • National Science Foundation grant #ACI008915 and #CCR0092065 • NIH Insight Project University of Utah