Fast computation of voxel-level brain connectivity maps from resting-state functional MRI using l1-norm as approximation of Pearson's temporal correlation: Proof-of-concept and example vector hardware implementation

An outstanding issue in graph-based analysis of resting-state functional MRI is choice of network nodes. Individual consideration of entire brain voxels may represent a less biased approach than parcellating the cortex according to pre-determined atlases, but entails establishing connectedness for 1(9)-1(11) links, with often prohibitive computational cost. Using a representative Human Connectome Project dataset, we show that, following appropriate time-series normalization, it may be possible to accelerate connectivity determination replacing Pearson correlation with l(1)-norm. Even though the adjacency matrices derived from correlation coefficients and l(1)-norms are not identical, their similarity is high. Further, we describe and provide in full an example vector hardware implementation of l(1)-norm on an array of 4096 zero instruction-set processors. Calculation times <1000 s are attainable, removing the major deterrent to voxel-based resting-sate network mapping and revealing fine-grained node degree heterogeneity. l(1) norm should be given consideration as a substitute for correlation in very high-density resting-state functional connectivity analyses. (C) 2014 IPEM. Published by Elsevier Ltd. All rights reserved.