The current high-resolution diffusion MRI (dMRI) protocol at 7 Tesla (7T) for the Human Connectome Project (HCP) uses single-channel RF transmission and is limited to a multiband (MB) factor of 2 (MB=2) because of SAR. The objective of this study was to investigate numerically the highest possible slice acceleration (or minimum volume TR) at 7T when designing parallel transmission (pTx) MB RF pulses to acquire HCP-type whole-brain dMRI with ~1-mm isotropic resolutions.We considered a 16-channel transmit, 32-channel receive array¹ specifically built for pTx and high-resolution neuroimaging at 7T. The transmit portion of this array, comprising 16 resonance elements arranged in multiple z-stacked rows, was modeled in a commercial FDTD software (XFDTD, Remcom, State college, PA, USA) and was loaded with a human tissue model (Duke, Virtual Family). Electromagnetic field of each element was simulated by driving one element at a time in the presence of all others. For more realistic simulations, a ΔB0 map in the brain was also synthetized using the method of Salomir et al², assuming second order shimming. Pulses were parametrized as spokes pulses whose number was varied from one (RF shimming) through three. The sub-pulses were Hanning-window apodized 3.4/1.7 ms sinc pulses with 1.5/1.85 kHz bandwidths for the 180° refocusing and 90° excitation pulses, respectively, hence with the same time-bandwidth products as in HCP³. The pulse design was conducted separately for the two types of pulses, and in both cases jointly on 20 sagittal 6 mm-thick slices covering the brain, thereby repeating each pulse 6 times to achieve whole brain coverage with 1-mm resolution⁴. The active-set optimization algorithm was employed to solve the pulse design optimization problem under explicit SAR constraints⁵. The peak-power constraint was not incorporated because it can be alleviated in a second step, e.g. using time-shifting⁶ and phase scrambling⁷ methods. Calculations were performed with Matlab (The Mathworks, Natick, MA, USA) and on a GPU card using CUDA. To spare the full Bloch integration of the time-discretized sub-pulses, we used the technique reported by Warren⁸ and derived from average Hamiltonian theory, which approximates the rotation matrix of a sub-pulse by simply concatenating three matrices. Pulses were designed by varying the volume TR, 7 s being the current value employed in the 7T HCP dMRI with MB=2. Pulse performance was evaluated using full Bloch simulations. For a given TR, the joint design of the 20 refocusing or excitation pTx single-band pulses was accomplished within 20, 42 and 80 seconds for 1, 2 and 3 spokes, respectively. The flip angle normalized root mean square error (NRMSE) vs. TR is shown in Fig. 1. At a TR of 3.5 s (corresponding to MB=4 and half the TR used in the current 7T HCP dMRI protocol), our pTx pulses improved the excitation fidelity (NRMSE reduced by up to 57% compared to a 7T CP mode). For TR=4.7 s (corresponding to MB=3), our pTx pulses further reduced excitation error by up to 72%, achieving better excitation fidelity than a 3T CP mode . At TR=7 s (as in the 7T HCP dMRI with MB=2), NRMSE was reduced by up to 82% in our design. Including ΔB0 in simulations led to different NRMSE vs TR tradeoffs compared to that obtained otherwise. Using Warren’s approximation introduced ~7% error for large ΔB0 values (Fig. 2). The refocusing pulses produced near-180° flip angles across the whole brain (Fig. 3), except for the regions of large ΔB0 (~300 Hz) due to limited pulse bandwidths. Similar performance was obtained for the excitation pulses.Our numerical results show that MB=4 in principle is possible when acquiring HCP-type dMRI data at 7T with ~1-mm isotropic resolutions. Compared to the current HCP protocol using MB=2, this doubled slice acceleration can translate into a twofold reduction in acquisition time for constant diffusion directions or twofold increase in the diffusion directions for constant acquisition time. Given the fact that most 7T pTx sites only have 8 RF channels, we repeated calculations for the same dMRI protocol using an 8-element head array and found that it is also possible to achieve MB=4 with an 8-channel coil, while improving excitation homogeneity compared to a CP mode; however the excitation fidelity is not as good as that of the 16-channel coil. Numerical results predict that parallel RF transmission can be used to achieve four-fold slice acceleration (corresponding to a volume TR as short as 3.5 seconds) for 7T HCP-type dMRI with ~1-mm isotropic resolutions, while improving flip angle homogeneity across the entire brain as compared to a single-channel CP mode.