Speeding up Mutual Information Computation Using NVIDIA CUDA Hardware

We present an efficient method for mutual information (MI) computation between images (2D or 3D) for NVIDIA's `compute unified device architecture' (CUDA) compatible devices. Efficient parallelization of MI is particularly challenging on a `graphics processor unit' (GPU) due to the need for histogram-based calculation of joint and marginal probability mass functions (pmfs) with large number of bins. The data-dependent (unpredictable) nature of the updates to the histogram, together with hardware limitations of the GPU (lack of synchronization primitives and limited memory caching mechanisms) can make GPU-based computation inefficient. To overcome these limitation, we approximate the pmfs, using a down-sampled version of the joint- histogram which avoids memory update problems. Our CUDA implementation improves the efficiency of MI calculations by a factor of 25 compared to a standard CPU- based implementation and can be used in MI-based image registration applications.