Comput. & Graphics Vol. 16, No. 1, pp. 85-100, 1992 0097-8493/92 $5.00 + .00 Printed in Great Britain, © 1992 Pergamon Press pie Technical Section VOLUME-PRIMITIVE BASED THREE-DIMENSIONAL MEDICAL IMAGE RENDERING: CUSTOMIZED ARCHITECTURAL APPROACHES MARTIN R. STYTZ* Air Force Institute of Technology, Department of Electrical and Computer Engineering, Wright-Patterson AFB, OH 45433 and OPHIR FRIEDER George Mason University, Department of Computer Science, Fairfax, VA 22030 Abstract--This paper examines seven computer architectures specificallydesigned to rapidly render 3D medical images from voxel data. The paper opens with a discussionof work on architectures for 3D medical image renderingand then specifiesparameters for assessing the performanceof a 3D medical image rendering architecture. We then describe and assessthe 3DP 4, the Cube, the INSIGHT, the PARCUM II, the PICAP II, the Voxel Flinger, and the Voxel Processor architectures. For each machine the rendering speed, image resolution, underlying data model, image quality, parallel processingstrategy, and 3D display technique are discussed. The architecture for each machine is characterized by its data storage technique, computational architecture, and parallelism strategy. l. INTRODUCTION 3D medical imaging is the subspecialty of volume vi- sualization and computer graphics that addresses issues associated with the rendering and display of medical image data.* 3D medical imaging had its beginnings in the late 1970s and has since become a recognized adjunct to traditional 2D medical data display tech- niques. 3D medical imaging is currently used to plan craniofacial surgery, for traumatology, in orthopedics, for disease diagnosis, for teaching, and for radiation treatment planning. To enable these uses, 3D medical imaging attempts to compute an accurate, clinically useful 3D depiction of the data as rapidly as possible. Early efforts in 3D medical imaging, see [ 1, 2 ], con- centrated upon demonstrating the clinical utility of 3D medical imaging and used low-power, general-purpose processors for the image rendering task. Their primary focus was upon reducing image rendering time by per- forming data classification in a preprocessing step, with the output of this step being a 1D or 2D data model of a particular organ. However, rendering still took hours, so batch processing was used to perform 3D medical image rendering. Subsequently, researchers investigated the use of customized architectures and alternative data models for reducing image rendering time and increasing the utility of the image by im- proving its quality and information content. Before turning to a description of the customized 3D medical image rendering machines that are the focus of the present paper, a brief review of other machines for 3D imaging is presented below. * To whom correspondence should be addressed. * The data to be imagedis gathered usinga medicalimaging modality such as Positron Emission Tomography, Comput- erized Tomography, Magnetic Resonance Imaging, Ultra- sound, or Single Photon Emission Computed Tomography. The MIPG machines*[ 1, 3-6 ] and work at Cornell, [2], led the way with their demonstration of the utility of 3D medical images constructed from extracted con- tour or surface models of individual organs.* The utility of the data models, surface extraction techniques, and shading algorithms employed in these machines is demonstrated by their use in later machines. Farrell's machine[9 ] demonstrates the usefulness of false color, composited images in a medical imaging environment. The two Fuchs/Poulton machines[10, 11] establish that the pixel coloring computational bottleneck can be overcome by expending hardware resources at the pixel level. The Fuchs/Poulton machines also dem- onstrate the usefulness of a pipeline architecture for computing volume visualizations. The work at the Mayo Clinic, described in [ 12-14 ], demonstrates techniques for rendering images with a set of cooper- ating processes. The rendering tool is named the AN- ALYZE* system and has a capability for the display of both 2D contour-based and 3D volume-based im- ages. Several current commercial machines reflect a nar- rower range of approaches to the 3D medical image rendering problem than is the case for the research machines. The Ardent Titan,* AT&T Pixel Machine, the Pixar machines, the Stellar GS2000 t, and the Sun t Originallyat the State University of New York in Buffalo, later at the University of Pennsylvania. *A MIPG designed, PC-based rendering system using al- gorithms derived from the early MIPG work is described in [7,8]. t ANALYZE for UNIX workstationsis available from CE- MAX, Inc, 46750 Freemont Blvd., Suite 207, Freemont, CA, 94538. *The Stellar and Ardent Computer Corporations merged to form Stardent Computer in 1989. The Ardent architecture is marketed as the Stardent 1000 and 3000 machines, the Stellar architecture is marketed as the Stardent 2000 machine. 85