High Performance Medical Image Registration Using a Distributed Blackboard Architecture

Abstract

A major drawback of medical image registration techniques is the performance bottleneck associated with similarity computation. Such bottlenecks limit registration applications in situations where fast execution times are required. In this paper a novel framework for high performance intensity-based medical image registration is presented. Geometric alignment of both reference and sensed images is achieved through a combination of scaling, translation, and rotation. Crucially, similarity computation is performed intelligently by knowledge sources (KSs) organised in a worker/manager model. The KSs work in parallel and communicate with each other by means of a distributed blackboard architecture. Partitioning of the blackboard is used to balance communication and processing workloads. The registration framework presented demonstrates the flexibility of the coarse-grained parallelism employed and shows how high performance medical image registration can be achieved with non-specialised architectures. Experimental results obtained during testing show that substantial speedups can be achieved