Performance of 3D Deconvolution Algorithms on Multi-Core and Many-Core Architectures

Deconvolution algorithms are commonly used to remove optical distortion from fluorescence microscopy images. Many such algorithms have been proposed, but those that produce the best image restoration results are iterative. Typically, each iteration involves one or more 3D convolutions, resulting in execution times of tens of seconds to several minutes for common image sizes on single-core computers. Fortunately, most of the constituent computational primitives in deconvolution algorithms are readily parallelized on shared memory architectures. In this paper, we analyze the performance of three deconvolution algorithms implemented on modern multi-core central processing units and on many-core graphics processing units. We discuss the computational primitives in the deconvolution algorithms and their implementations, and compare performance of the two implementations on recent parallel processing architectures.