A GPU-accelerated Framework for Fast Mapping of Dense Functional Connectomes

摘要

In the context of voxel-based modalities like functional magnetic resonance imaging (fMRI), a dense connectome can be treated as a large-scale network where single voxels are directly used to define brain network nodes. Contrary to parcellated connectomes, dense connectomes have higher spatial resolution and are immune from the parcellation quality. However, the analysis of dense connectomes basically requires more powerful computing and storage capacities. Here, we proposed a graphics processing unit (GPU)-accelerated framework to perform fast mapping of dense functional connectomes. Specifically, the framework is scalable to high voxel-resolution imaging data (<2mm) and can construct large-scale functional brain networks with lower time and memory overheads. Based on the proposed framework, three functional connectivity measures (Pearson's, Spearman's and Kendall's) were accelerated on the GPU for fast detection of possible functional links in dense connectomes. Experimental results demonstrated that our GPU acceleration for the Kendall's measure delivered a >50x speedup against both multi-core CPUs implementations and GPU-based related works.