Vis3D+ A tightly integrated GPU-accelerated computation and rendering framework for interactive 3D image visualization

Irfa Nisar

doi:10.32920/ryerson.14668293.v1

Abstract

This thesis presents extensions to an interactive 3D image visualization framework. The existing software framework provides functionality for interactively visualizing 3D medical data. The extensions consist of software modules that execute directly on the graphics hardware, utilizing the massively parallel, general-purpose computing platform provided by modern graphics processing units (GPUs). These GPUbased software modules are designed to support the execution of volume image processing algorithms, implemented using recently available GPU programs known as “compute shaders”, as well as to support interactive editing of the algorithms’ output. The new modules are seamlessly integrated as new stages in a GPU-based rendering pipeline provided by the existing framework. In this thesis, an example volume image processing algorithm known as level set segmentation is implemented and demonstrated. In addition, a new editing module is demonstrated that enables user modification of this algorithm’s output by extending a pre-existing volume “painting” interface.

Highlights

Medical imaging currently plays a critical and expanding role in a host of clinical applications, from disease diagnosis and subsequent treatment planning, to surgical and radiotherapy planning, and even to intraoperative surgical navigation
graphics processing units (GPUs) were initially designed to efficiently render surfaces consisting of millions of polygons by taking advantage of the inherently parallel nature of polygon vertex and pixel operations, recent generations of graphics cards can be used as general-purpose parallel computing platforms and support programming for the GPU using high-level programming languages
Since this thesis is primarily concerned with the compute shader implementation and integration of volume image processing algorithms into the existing framework, we focus on demonstrating a working segmentation algorithm and some measurements of its performance, rather than on a formal analysis of segmentation accuracy and efficiency

Summary

Introduction

Medical imaging currently plays a critical and expanding role in a host of clinical applications, from disease diagnosis and subsequent treatment planning, to surgical and radiotherapy planning, and even to intraoperative surgical navigation. GPUs were initially designed to efficiently render surfaces consisting of millions of polygons by taking advantage of the inherently parallel nature of polygon vertex and pixel operations, recent generations of graphics cards can be used as general-purpose parallel computing platforms and support programming for the GPU using high-level programming languages. The result of these advances in GPU programmability is the ability to perform real-time surface rendering and real-time volume (image) rendering. Many volume image processing algorithms are data-parallel and require repeating operations on individual voxels or on a small local neighborhood of voxels

Objectives

Methods

Results

Conclusion