George J. Grevera, Jayaram K. Udupa
Medical Image Processing Group (MIPG)
Department of Radiology
423 Guardian Drive
University of Pennsylvania Health System
Philadelphia, PA 19104
voice: (215) 662-6780
fax: (215) 898-9145
email: grevera@mipg.upenn.edu, jay@mipg.upenn.edu

3. Presentation: Poster

4. Principal Author's Bio

George Grevera received a Ph.D. in Computer Science from the University of Pennsylvania, an M.S. in Electrical Engineering from Drexel University, and a B.S. in Computer Science and Biology Research (double major, Cum Laude) from the University of Scranton. He is currently a Research Assistant Professor in the Medical Image Processing Group at the University of Pennsylvania. His current research interests are medical visualization and registration.

5. Abstract

Abstract Text (250 words)

We describe a generalization of the ubiquitous Maximum Intensity Projection (MIP) method of volume visualization, which we name Generalized MIP (GMIP). Briefly, this new technique allows the user to specify a range of intensity values of interest within which the intensities are subjected to a transformation to enhance the object of interest and then a MIP is done on the transformed intensities. Whereas the MIP method projects the maximum intensity value along a particular projection line, the GMIP ignores any intensity values outside the specified range, and projects only the maximum of the transformed intensity value along a particular projection line. GMIP is computationally very efficient and is conceptually a simple extension of the MIP method but provides a great deal of new capabilities. For example, in CT, bone can obscure vasculature in MIP renditions, but this effect can be reduced by the GMIP method. GMIP also allows rendering soft tissue and even dark appearing structures in CT and MRI. We compare and contrast this new method with other well known methods such as surface and volume rendering.

Five Paragraph Summary

Purpose We describe an extension to the familiar MIP algorithm that we call the GMIP (Generalized Maximum Intensity Projection) method. It combines the classification step employed by volume and surface rendering with the maximum intensity projection of the MIP algorithm. We compare and contrast this new algorithm with traditional surface and volume rendering methods on a variety of data sets and demonstrate how useful visualization of almost any structure can be efficiently created without explicit hard or fuzzy segmentation.

Methods Given a 3D grey scene, traditional MIP projects the maximum grey value along lines cast through the data volume from a particular viewing direction. High intensity values therefore tend to dominate the resulting projection regardless of surrounding values (or other values along the line of projection) or distance of the high density value from the viewpoint (i.e., distance along the ray). To introduce more subtle detail, it has been suggested that not only should the maximum value be projected along a given ray but the distance from the viewpoint to the value should influence the projected value as well. MIP implementations tend to be computationally efficient and achieve interactive frame rates. Because of this, the data may be quickly and easily viewed from a variety of viewpoints. Therefore values that may not be maximal in one viewing direction may be maximal in another. In this manner, much of the entire volume of data may be inspected.

Surface rendering (SR), on the other hand, typically requires that a binary classification step occur before rendering. Many segmentation methods from fully manual to fully automatic have been proposed. After the classification step occurs, the surface of the object is then constructed by tiling it with some geometric primitive (triangles, voxels, or voxel faces). These geometric primitives are then rendered typically using relatively inexpensive dedicated hardware (in the form of graphics cards whose development has been spurred on by the computer games industry) and also by software-only techniques which are often faster than the former. Interactive frame rates can be achieved but the hardware can be overwhelmed by high resolution data (such as spiral CT data). But the main problem with this method lies in the classification and tiling steps which must be repeated if the desired object of interest hasn't been accurately segmented. And regardless of viewpoint, only surface data may be viewed.

Volume rendering (VR) algorithms extend the MIP algorithm by incorporating a "fuzzy" classification method in the form of a transfer function. This function is simply a table that contains and (red,green,blue,opacity) quadruple entry for each grey value. But rather than projecting the maximal table entry corresponding to the grey values occurring along each ray as in MIP, values are composited (blended) along the ray using the opacity values. The result is that nearer, more opaque values dominate the projected image. VR algorithms require much more computation than MIP. To achieve interactive frame rates, special purpose hardware has been developed. This hardware is expensive and is not flexible from a programming standpoint as general purpose computational platforms. But if this hardware is available, interactive frame rates may be achieved. Furthermore, in contrast with SR, the classification function may be changed at interactive rates as well.

Hybrid methods, combining SR and VR, have been proposed as well. One such method, developed by our group, is shell rendering which combines SR and VR by representing the surface as a fuzzy shell. At one extreme, shell rendering behaves like SR if the user specifies a shell that is only one voxel thick. At the other extreme, shell rendering behaves like VR when the user specifies a shell that is as thick as the entire data set. The idea behind shell rendering is that the user can easily indicate the approximate surface of the object of interest. This surface may then be rendered using general purpose computers at interactive frame rates using very large data sets. The shell idea can also perform MIP rendering.

GMIP is a generalization of MIP which incorporates a classification step similar to SR, VR, or shell rendering followed by a subsequent maximum intensity projection. GMIP classification may employ a transfer function as in VR which simply maps grey values to other grey values. Additionally, the GMIP may also employ segmentations from SR (or the shell definition from shell rendering) as a mask in which the MIP is restricted. Since compositing is not employed in the GMIP as in VR, special hardware is not required to achieve interactive frame rates. If hand-segmented data sets are available from prior surface renderings, they may be used by the GMIP as well. Similarly, the GMIP may incorporate shells as well. As in MIP, GMIP does not require explicit hard or fuzzy segmentation. The transfer functions for GMIP do not have the restrictions coming from tissue fractions as in VR, and they can simultaneously render multiple objects of disparate intensity characteristics.

Results We compare and contrast the new GMIP algorithm with other popular rendering algorithms such as traditional MIP, surface rendering, and volume rendering with renditions of a wide variety of medical image data sets. We also demonstrate how, without explicit segmentation, multiple objects of disparate intensity characteristics can be simultaneously rendered via GMIP for a full free-form browsing of the scene.

New Work to be Presented Although the MIP is a very useful and traditional algorithm, the GMIP is a new algorithm that extends the usefulness of MIP while maintaining its efficient computation and conceptual simplicity. GMIP extends all advantages of MIP from the brightest objects to multiple objects of any intensity characteristics.

Conclusions The GMIP is a useful, efficient rendering algorithm that can be effectively used to replace traditional MIP and as an adjunct to other rendering algorithms.

6. Supplemental Information: None.

7. Keywords: surface rendering, volume rendering, maximum intensity projection