Distributed Medical Volume Registration

Introduction

The ability to visualise hidden structures in detail using 3-D volume data has become a valuable resource in medical imaging applications (Maintz & Viergever, 1998). Importantly, the alignment of volumes enables the combination of different structural and functional information for diagnosis and planning purposes (Pluim, Maintz, & Viergever, 2003). Transform optimisation, re-sampling, and similarity calculation form the basic stages of a registration process (Zitova & Flusser, 2003). During transform optimisation, translation and rotation parameters which geometrically map points in the reference (fixed) volume to points in the sensed (moving) volume are estimated. Once estimated, voxel intensities which are mapped into non-discrete co-ordinates are interpolated during the re-sampling stage. After re-sampling, a metric is used for similarity calculation in which the degree of likeness between corresponding volumes is evaluated. Optimisation of the similarity measure is the goal of the registration process and is achieved by seeking the best transform. All possible transform parameters therefore define the search space. Due to the iterative nature of registration algorithms similarity calculation represent a considerable performance bottleneck which limits the speed of time critical clinical applications.

The use of parallel computing to overcome these time constraints has become an important research area (Nicolescu & Jonker, 2000). Conveniently, many of the similarity calculation strategies employed in medical registration are inherently parallel and therefore well suited to distribution. An important consideration when adopting a parallel processing approach is the architecture of the host system. In a computer constructed of multiple processors with shared-memory, data distribution is not required. These systems are viewed as tightly-coupled architectures. In contrast, a loosely-coupled architecture consists of multiple computers in different locations. Loosely-coupled architectures therefore require data distribution, communication, and accumulation mechanisms. Logically, the most effective distribution scheme will depend on the architecture of the host system (Seinstra, Koelma, & Geusebroek, 2002). The two contrasting architectures of host systems are illustrated in Figure 1.

Figure 1.

Tightly vs. loosely-coupled architectures. Data is either fetched from main memory via a memory bus, or is transferred over a communications network

Background

In the context of parallel processing, registration of medical data has been achieved by Warfield et al. (Warfield, Jolesz, and Kikinis, 1998) who introduced a non-rigid algorithm based on the work-pile paradigm. Their goal was to develop an inter-patient registration algorithm which can be applied without operator intervention, to a database of several hundred scans. In an initial step, each scan is segmented using a statistical classification method. This pre-processing stage is used to identify different tissue types including skin, white matter, grey matter, and bone structure. Once segmented, a transform which brings these features into alignment is estimated. The system employs a message passing interface and cluster of symmetric multi-processors to execute parallel similarity calculation operations using multiple threads. Crucially, work is dynamically load balanced. Results published by the group show that successful registration of 256 × 256 × 52 voxel brain scans can been achieved in minutes rather than hours.