Simultaneous Multi-Slice (SMS) methods [1] accelerate the temporal acquisition rate of EPI acquisitions. SMS-EPI is gaining wide acceptance in neuroimaging applications because it can improve imaging efficiency with limited SNR penalty. Like all EPI-based methods, however, SMS-EPI is limited by inherent imaging artifacts. Correcting Nyquist ghosts in SMS-EPI is particularly challenging because different slices often have different EPI phase errors [2], necessitating slice-specific ghost correction at the slice separation stage [3, 4]. Conventional ghost correction approaches for SMS data employ navigator signals acquired from the entire collapsed slice-group, making identification of phase errors associated with each individual slice difficult. This work describes an approach to correct slice-specific phase errors in SMS-EPI data through the Dual-Polarity GRAPPA (DPG) [5] method. We demonstrate that DPG can reduce Nyquist ghosting in SMS data compared to the combined even-odd slice-GRAPPA and slice-specific ghost correction in [4], particularly when individual slices introduce non-linear phase errors to the collapsed slice-group.

SMS-EPI data were acquired using a standard Blipped-CAIPI pulse sequence [6]. In vivo data were acquired using a 2.5mm isotropic single-shot EPI protocol with TR=1 s, TE=28 ms, 80×80 matrix, 9 slices, MultiBand (MB) factor 3, no FOV shifting, and no in-plane acceleration (R=1). Phantom data were acquired using a 2.6mm isotropic protocol with TR=1 s, TE=27 ms, 76×76 matrix, 9 slices, MB factor 3, FOV/3 CAIPI shift, and in-plane acceleration of R=2. All data were collected on a Siemens 7T whole-body scanner equipped with SC72 body gradients and a custom-made 32-channel brain array receive coil [7]. Images were reconstructed using MATLAB running on a Linux workstation and were compared with default online reconstructions.

Two temporally encoded frames of calibration data were acquired for each slice, with reversed readout gradient polarity, using FLEET [8] to limit temporal phase errors. The calibration data were then sorted by readout gradient polarity, RO⁺ versus RO⁻, and a set of ghost-free “target” data were generated using GESTE, as in [5]. Next, calibration “source” data were generated by synthesizing an SMS-EPI slice-group acquisition from the single-slice calibration data, including appropriate Blipped-CAIPI phase shifting. This process mirrors conventional slice-GRAPPA training [6], but here yields two sets of source data from the temporally encoded data: each consisting of RO⁺ or RO⁻ data only. DPG kernel coefficients were computed from the synthesized collapsed-slice calibration source data and a single uncollapsed slice from the ghost-free target data. This calibration process was repeated to estimate different DPG coefficients for each individual slice in the SMS slice group.

Reconstruction of each acquired collapsed slice group was performed by (a) removing the absolute phase imposed on the slice group due to CAIPI blipping, according to the position of the slices relative to the scanner isocenter, then (b) convolving appropriate DPG kernels with the data to yield ghost-corrected images for each slice in the slice group. For the R=2 data, a single DPG kernel was used to perform slice separation and data recovery simultaneously. Coil images were combined using root-sum-of-squares (RSOS). For comparison, current state-of-the-art images were generated by the on-scanner SMS-EPI reconstruction program.

Conventional Nyquist ghost correction (NGC) seeks to estimate linear and constant phase errors from the difference in phase between navigator signals measured prior to the EPI echo train. Fig. 1 illustrates these phase difference signals for three uncollapsed slices of the in vivo calibration data and their corresponding collapsed slice group. Two slices (9 and 6) exhibit the expected linear character, whereas the third (Slice 3) does not. In the NGC signal for the slice group, bottom of Fig. 1, noise is largest in regions corresponding to the Slice 3 non-linearity. DPG is able to correct non-linear phase effects [5] and provide slice-specific ghost correction, yielding images from SMS-EPI data with significantly lower levels of ghosting as shown in Fig. 2 and highlighted by the arrows. Phantom images in Fig. 3 demonstrate that DPG can concurrently perform data recovery in SMS-EPI data acquired using in-plane acceleration. Although ghosting levels in the on-line and DPG phantom images are comparable, the DPG-SMS images show lower levels of phase interference artifacts in the upper slices, as highlighted by the arrows.

Our results demonstrate that DPG can simultaneously perform data recovery, slice separation, and ghost correction on SMS-EPI data with both slice and in-plane accelerations. Its effectiveness may be limited by the size of the DPG kernels, however, and depend on the MB factor and in-plane acceleration rate. Understanding the interaction between kernel size, MB factor, and in-plane acceleration rate is a focus of current research.