Three-dimensional medical image analysis using local dynamic algorithm selection on a multiple-instruction, multiple-data architecture.

Abstract

The dissertation outlines development of a medical imaging machine which renders 3D images from voxel data within a MIMD multiprocessor architecture at interactive rates. Interactive performance is achieved using local dynamic selection of the optimum adaptive recursive hidden-surface removal algorithm. A survey of the medical imaging, graphics, and medical imaging modality literature is provided. A description of Computerized Tomography, Magnetic Resonance Imaging, Positron Emission Tomography, Single Photon Emission Computed Tomography, and Ultrasound imaging modalities is presented. Previous work in 3D volume rendering graphics techniques and data models is introduced. Eleven medical imaging machines are examined with emphasis on characterization of the major innovation(s) and performance of each machine. A five stage image processing pipeline is described. The front-end of the pipeline performs user interface and volume rendering operations in software upon voxel data. The back-end of the pipeline accomplishes pixel shading in hardware. The recursive hidden-surface removal operation in the front-end of the pipeline is found to be the pipeline bottleneck. To provide a framework for development of improved recursive hidden-surface removal algorithms, a definition of recursive hidden-surface removal algorithms is formulated. Back-to-front and front-to-back recursive hidden-surface removal algorithms suitable for scene editing are developed from this definition and their performance is investigated. A processing termination capability for back-to-front hidden-surface removal algorithms is developed from the definition, the resulting adaptive back-to-front recursive hidden-surface removal algorithm is described and its performance is examined. A processing termination capability for front-to-back hidden-surface removal algorithms is developed from the definition, the resulting adaptive front-to-back recursive hidden-surface removal algorithm is specified and its performance is investigated. An algorithm to accomplish individul-processor-based dynamic selection of either the adaptive back-to-front or adaptive front-to-back hidden-surface removal algorithm based on the processing situation faced by each processor is described. Dynamic algorithm selection is shown to reduce the amount of time spent rendering the image by allowing each processor to employ the adaptive hidden-surface removal algorithm with minimum local image rendering time.