MR physicists and Neuroscientists interested in ultra-fast, state-of-the-art functional MRI at 3T.To compare two slice-accelerated EPI sequences using controlled aliasing (simultaneous-multi-slice vs. segmented 3D-EPI) for ultra-fast functional MRI under realistic conditions at 3T.

A state-of-the-art simultaneous multi-slice (SMS) blipped-CAIPI EPI sequence^1,2 is compared to 3D-EPI³ featuring 2D-CAIPIRINHA⁴ sampling^5,6, optimized for ultra-fast fMRI at 3T (Siemens Prisma, 64-channel head coil). Although both acquisition strategies for controlled aliasing have been shown to be equivalent⁷ and can be reconstructed the same way with identical g-factor penalties⁶, a stringent comparison is not as trivial practice: different reconstruction algorithms are typically chosen for the two, and they are differently affected by physiological noise. In this work, individually optimized 3D-EPI and SMS-EPI sequences and dedicated image reconstructions but otherwise matched parameters, were therefore utilized, under the constraint of 2.4mm isotropic axial whole-brain coverage in equal TR. The latter was selected short enough to ensure separation of aliased cardiac peaks from BOLD signal in the frequency domain. Utilizing a readout bandwidth of 2470Hz/pixel and reasonable acceleration factors^2,5,6 with both sequences a TR of 580ms at TE=30ms could be achieved.

While only CHESS⁸-based fat-saturation is feasible with the SMS-EPI sequence, one can make use of bipolar binomial-11 water excitation⁹ in case of 3D-EPI, which was previously shown to have two beneficial side effects¹⁰: reduction of the TR (save ~12ms per excitation), and increase of gray matter SNR (avoiding unintentional magnetization transfer contrast). A smaller total acceleration factor R=6 could therefore be used with 3D-EPI (3x2⁽¹⁾ 2D-CAIPIRINHA sampling⁴) compared to SMS-EPI, which requires a slice acceleration factor of MB=8 (FoV/3 slice shift) to achieve the same TR=580ms. The former employed vendor-provided 2D-CAIPIRINHA reconstruction (“IcePAT”) and the latter dedicated Slice-GRAPPA reconstruction¹, both implemented on the scanner.

EXPERIMENTS: Five subjects underwent short resting state scans with both sequences (830 volumes in 8 minutes). Additionally, a conventional, singleband 2D-EPI (3mm isotropic, 32 slices + 25% gap, TR=2s, 240 volumes) and a 1mm isotropic T1-weighted anatomical scan were acquired. The chronological scanning order was altered pseudo-randomly between subjects. Ernst angles of 84°/47°/16° (2D-EPI/SMS-EPI/3D-EPI) according to T1=1500ms were used for excitation.

PROCESSING: The processing pipeline consisted of: (1) removal of the initial 10s of data, (2) motion correction to the first retained volume, and (3) temporal filtering prior to (4) computation of temporal SNR (mean/standard deviation with respect to time) and sensitivity per unit scan time (SEN:=SNR/√TR). Two alternative temporal filters were applied: a highpass (0.01Hz cut-off for detrending only) and a bandpass (0.01Hz and 0.1Hz cut-off to additionally remove physiological noise)¹¹. Following the latter computation of sensitivity maps is not applicable since the effective temporal resolutions have already been equalized. Finally, the mean, SNR and SEN maps were normalized to the 2mm MNI template in order to compute the average over all subjects. Additionally, whole-brain temporal spectra were computed and averaged over all subjects.

Fig. 1 (left) shows three axial example slices of the group-averaged mean maps for 2D-EPI, SMS-EPI and 3D-EPI. The sampling pattern depicted at the bottom right indicates the 2D-CAIPIRINHA trajectory used for 3D-EPI. Temporal spectra of the highpass-filtered data (right, top) demonstrate a clear separation of the respiratory peak and the first alias of the cardiac peak from BOLD signal with SMS-EPI and 3D-EPI as opposed to 2D-EPI. Fig. 2 shows corresponding sensitivity and temporal SNR maps for the highpass-filtered and bandpass-filtered data, respectively. Both SEN and SNR of 3D-EPI appear to be superior to 2D-EPI throughout the brain (despite higher resolution), while SMS-EPI suffers from SEN and SNR drops, most apparent in the center of the brain.The temporal spectra demonstrate that physiological signals, which usually “pollute” typical TR=2s EPI data, are clearly separable by using ultra-fast EPIs. Accordingly, the bandpass-filtered SNR map of 2D-EPI still contains a lot of “false signal”, whereas fair comparison of the fast acquisitions is applicable. We observe a clear SNR advantage throughout the brain with the less undersampled 3D-EPI (R=6) over SMS-EPI (MB=8). On the other hand, comparison of the SNR and SEN maps indicates a slightly more positive effect of excluding high-frequency components for SMS-EPI over 3D-EPI. Advanced methods for physiological noise removal, such as RETROICOR¹² and RVHRCOR¹³, may even improve SNR, in particular for 3D-EPI¹⁴.The present comparison of sequences for ultra-fast whole-brain fMRI at 3T shows that 3D-(CAIPIRINHA-)EPI can have clear benefits over 2D-(SMS-)EPI at identical spatial and temporal resolution. With regard to the separation of physiological noise from BOLD signal, both sequences are favorable over typically “slow” singleband 2D-EPI.