Slice-accelerated EPI using multiband RF pulses that allow simultaneous multi-slice (SMS) acquisition of BOLD contrast images^1,2 can significantly enhance the temporal and spatial resolution of fMRI. This advance enables whole-brain fMRI at sub-second volume-TRs,^3,4 which is advantageous for resting-state fMRI, which attempts to capture sub-second frequency fluctuations.^5,6 However, the computational processes required for slice-unaliasing can be taxing for the reconstruction computer, so data can be observed far from real-time, and reconstruction may lag up to tens of minutes behind the acquisition. A channel compression algorithm has been proposed to counter these computational demands.⁷ Here, functional networks derived from resting-state scans are studied as a function of slice acceleration, slice-GRAPPA kernel size,² and channel compression to find an optimal solution in terms of data quality and reconstruction time for a conventional 3.0 T scanner.All measurements were performed using a Siemens Tim Trio 3.0 T MRI scanner equipped with the fastest Step4 image reconstruction computer, and using the vendor-supplied 32-channel head coil. Slice-accelerated BOLD images were acquired using Siemens WIP770A sequence with modified reconstruction code that improved computational speed and enabled channel compression. All resting-state scans employed ~6 min of data acquisition while subjects rested with eyes open. A 3mm isotropic, unaccelerated scan was acquired with the conventional BOLD sequence as a reference point. Slice-accelerated scans were acquired at 2mm nominal resolution (6/8 partial-fourier), with SMS=1, 3, 4, 6 or 8, enabling TR=5.90, 2.00, 1.44, 0.93 and 0.76 sec respectively. TE=30 ms, BW~1600 Hz/px, echo spacing~0.71 ms for all 2mm scans. The raw data was reconstructed at the scanner multiple times with varying slice-GRAPPA kernel sizes (3×3, 5×5, 7×7) and channel compression factors (CCR=1, 2, 4). Notation throughout indicates kernel size and CCR as: kX×kY×CCR. Seed-based functional connectivity analysis was performed using in-house scripts.⁸ Slice-time correction was performed for the 3mm and MB1 scans, but omitted for all slice-accelerated scans.Fig. 1 shows an example for the Default Network (posterior cingulate seed⁸) where SMS=4 shows a larger number of voxels detected than the SMS=8 case, using the same reconstruction parameters. For a given SMS factor, larger kernel sizes with CCR=1 or 2 yielded improved tSNR over the default 3×3×1 reconstruction, however this generally did not improve network definition. Fig. 2 shows example maps of the Default Network for SMS=8, with different levels of channel compression. CCR=2 has minimal effect on network definition, but CCR=4 resulted in lower tSNR for all kernel sizes, and significant loss of network definition as well as introducing significant noise and false positives. Similar results are seen in the motor and visual networks. Fig. 3 highlights differences between pairs of reconstructions (from the same acquisition). z scores for major network nodes are slightly lower when CCR>1, usually ~0.05 – these have been masked out. Fig. 4 shows voxel counts for Default and motor networks as a function of SMS and reconstruction parameters. While SMS=8 (and maximum number of time points) aids motor network definition, the default network is optimally shown with SMS=4.Despite the increase in the number of time-points acquired, increasing the SMS factor to 6 or 8 was not always beneficial for resting-state analysis. While slice-accelerated scans always resulted in better network definition over unaccelerated scans, for some networks, SMS=4 provided the maximum number of voxels above threshold. Motor and visual networks benefited the most from SMS=8 acceleration. For a given SMS-acquired dataset, 5×5 kernels had been commonly used for SMS>6 accelerations,^2,3 but this is unnecessary for fcMRI analysis. Although the tSNR is increased by the larger kernel size (after normalizing out other effects), this had minimal effect on network definition, while reconstruction times were heavily impacted. CCR=2 had minimal effect on data from SMS=4 and SMS=8 scans reconstructed with any kernel size, however CCR=4 introduced unacceptable levels of false activations and loss of network definition. Without channel compression, increasing the kernel size lengthened the reconstruction time by ~40 %. CCR=2 results in a 10% time reduction for a 3×3 kernel, but ~40% time reduction for 5×5. One acceptable compromise for maximum tSNR without impacting reconstruction time may be to use 5×5×2, although similar fcMRI maps will likely be achievable with 3×3×1. These conclusions relate specifically to the use of the Siemens 32-channel head coil on a Tim Trio scanner – higher kernel sizes and compression factors may be optimal with higher-array head coils or other scanner platforms, especially those equipped with GPU-assisted reconstruction computers. Future work will consider the effect of physiological noise effects over the different TR values on the results above.