Importance Driven Volume Rendering Authors: I. Viola, A. Kanitsar, M. Gröler Visualization II Instructor: Jessica Crouch Problem • Volume visualization has inherent difficulties with occlusion – Exterior surfaces occlude interior surfaces – Interior surfaces might be more interesting or important • How do you render a volume so that you see the most important parts of the data? Motivation • 3D medical image visualization in general • Specific example: liver tumor visualization – For radiotherapy planning – For surgery planning • Need to clearly see: – Tumor location, size, shape – Blood vessel tree – Parenchyma http://www.vislab.uq.edu.au/research/liver/images/3D_liver_model.jpg (functional tissue, as opposed to connective tissue) http://www.anatomyatlases.org/atlasofanatomy/plate33/images/33-4_static.jpg http://mbi.dkfz-heidelberg.de/projects/liver/images/vessels.jpg Methods • Approaches add visual emphasis to important portions of the scene – How do you define importance? – How do you add visual emphasis? Previous Work: Transfer Functions • Specify a function (probably non-linear) that maps scalar voxel data to color/opacity – Simplest example: threshold function – Others? • Map data through transfer function, then use a standard volume rendering algorithm – MIP (Maximum Intensity Projection) – Ray casting – Volume Splatting Previous Work: Transfer Functions • Assumes the important parts of the data can be specified using ranges of voxel intensities • How could this scheme work for the liver? Previous Work: Focus + Context Rendering • Focus: – To which part of the scene is the viewer paying attention? • Eye tracking – Render that part with more detail, with magnification, or other enhancement – Let focus define importance • Context: – The relationship (distance) between a part of the scene and the focus (focal) point – Detail or magnification gradually fades as distance increases Previous Work: Focus + Context Rendering • Has the unique property that the viewer determines what part of the visualization is emphasized – When is this appropriate? – How would this work for the liver application? Previous Work: Methods for Adding Visual Emphasis • Sparse Representation: – Just render important parts of the data • Iso-surfaces • Contour lines / curves (approximation of wireframe) • Cut-Away Views: – Traditional artist rendering of anatomy, layered structures – Clipping plane(s) or curved clipping surface(s) Importance Driven Rendering • Basic idea: – Wherever a less important object occludes a more important object, render the less important object sparsely so that the more important object is mostly visible. • Assumes importance is user-defined & static • View dependent approach – Not focus dependent Necessary Model Data • Each voxel is assigned to an object – Segmentation, classification • Each object is assigned a level of importance – In addition to color, opacity Rendering Logic • Sparseness level of a surface patch indicates what percentage of its footprint on the image plane it colors • Sparseness level is assigned to a surface patch based on whether it is more or less important than other surfaces it occludes (or is occluded by) Importance Compositing • Option 1: Maximum Importance Projection – For each pixel, render the most important voxel that projects onto the pixel – Assign sparseness level of 0 or 1 to each point 2 3 1 Importance Compositing: MImP • Efficient, fast? • How depth perception work? – Occlusion is an important depth cue • Naïve method of MImP: MImP Improved • Apply a conical rather than cylindrical cut-out to the objects surrounding an important object – Gradual change in rendering depth – Allows you to see cut away layers of surrounding objects, rather than a deep “hole” MImP Conical Cut-Out Implementation 1. Find the important object’s footprint on the image plane. 2. For each pixel in the footprint, find the maximum depth of all the object’s voxels that project to the pixel - Gives objects “back wall” 3. Create a conical clipping surface by projecting the back wall toward the image plane - Assuming a ray casting approach, just start tracing clipped rays at the clipping surface rather than at the image plane Importance Compositing • Option 2: Average Importance Compositing – Compute sum of object importances along a ray – Each surface point intersected by the ray is assigned a sparseness level based on the ratio of its importance to the sum of all intersected surface importances Average Importance Compositing • Efficiency of implementation? – Use binary object footprints • Artifacts: jaggy transitions – Enlarge and smooth footprints Levels of Sparseness • Given a voxel’s level of sparseness, how should it be rendered? – With MImP, render as solid or not at all – With Avg. Importance Compositing, more choices • Color and Opacity Modulation • Screen-Door Transparency • Volume Thinning Color and Opacity Modulation • Increase the opacity and possibly saturation of a voxel with increasing importance – Is playing with the color confusing? Screen-Door Transparency • Sparse surfaces are rendered as the fibers in a quadrilateral mesh – Higher sparseness level results in larger (coarser) mesh weave • Advantages/Disadvantages? – Depth perception – Distraction Volume Thinning • Render objects using a set of iso-surfaces • The number of iso-surface voxels to render per object depends on the sparsity level of the object Evaluation • Show rendered results for the Gecko and Monster datasets • All rendering off-line • No quantitative evaluation • No demonstration for the motivating problem: liver Evaluation • Several methods were presented, it would be nice to know which work the best • Would require human perception experiments – Task: plan a cutting path for removal of a liver tumor that avoids major vessels • Try it using MImP, Opacity Modulation, Screen Door, and Volume Thinning • Do people perform faster or more reliably using one of the types? Your prediction? Evaluation • Second consideration: – Path planning is an interactive task – Would like to rotate camera & change viewpoint • Which methods might be implemented in real- time? Conclusion • Have not addressed half the problem: how to assign importance to objects – Application dependent, but need automated methods Conclusion • Interesting, but incomplete work – Should be evaluated, validated • Most interesting idea is MImP with conical cut-out – Don’t have to use ray casting, although the authors do – Consider volume splatting • Avg. Importance Compositing methods likely to remain off-line rendering options for the near future Discussion Questions Consider applying this to visualization of a building volume (3D blueprints). 1. How might you assign importance? 2. Efficiency considerations aside, which of the rendering methods would be most visually effective? Discussion Questions What if you had a dataset that included air flow velocity for the volume, and you were interested in understanding the operation of the circulation system? 1. How might you assign importance? 2. Efficiency considerations aside, which of the rendering methods would be most visually effective? Discussion Questions • Implementation: – Could any of the methods described be implemented using the traditional forward- rendering graphics pipeline? • How, or why not? – What about implementation with a shading language? • A Z-Buffer type of approach for importance sorting?
Pages to are hidden for
"Importance Driven Volume Rendering"Please download to view full document