Fig 1.
Sequence diagram for SMS acquisition using a segmented EPI with multiple contrasts.
Looping is done over MB slice groups first, before looping over the EPI segments, to prevent from inter-slice contrast variation. An interleaved acquisition scheme was chosen to reduce cross talk between MB slice groups. CAIPIRINHA shifts: The orange area in Gz indicates the phase difference between EPI segments resulting in a phase modulation along PE depending on the z-position.
Fig 2.
Segmentation of k-space into shorter echo trains allows a flexible timing scheme, short echo times and the acquisition of multiple contrasts even for signals with fast T2* decay.
The duration between the EPI segments, i.e. acquisitions of adjacent ky lines, equals the repetition time (TR). Slice group looping is done before the looping of the EPI segments.
Fig 3.
ROI1 represents a voxel dominated by intravascular signal.
The intravascular signal describes typical AIF curve characteristics (left column, slice 15). In addition, representative regions in muscle tissue were selected for dynamic signal curves in tissue (right column, slice 19: ROI2 (solid), ROI3 (dotted)). High resolution anatomical images (TSE, sagittal) were used for selection and positioning of the ROIs. The diagrams in the left column show the signal dynamics in a major vessel for all three acquired contrasts (middle row) as well as for the separated signal components (bottom row). In the right column, the signals in tissue ROIs are displayed. For ROI2 (solid) and ROI3 (dotted) the three acquired contrasts (TE1, TE2, TE3) and the extrapolated signal for TE = 0 ms are depicted (middle row). The changes in T2* are shown in the bottom row. The sensitivity of T2* to inflow effects of CA leads to a decreased signal, which recovers to a constant level when CA is distributed in the ROI.
Fig 4.
SMS image reconstruction of the complete volume (24 slices, MB = 4, FOV/4 shift) performed with SG [17].
For kernel training measured SMS and SB data with identical imaging parameters were used. Four representative slices from one MB slice group are selected for further illustration (green box). A single slice with a TE1 = 9 ms of the reconstructed SMS image (center) and the reconstruction of the SB reference data (right) are shown in zoomed view. The results for the other echoes, TE2 and TE3, are depicted below the TE1 images, accordingly.
Fig 5.
The normalized subtraction maps of the different reconstruction methods are shown.
From left to right: Method 1, where the SG algorithm with measured data for ACS source (MB = 4) and ACS target (SB) was used for reconstruction. Method 2 (center), where the reconstruction is performed with synthesized ACS source data and, for comparison, a reconstruction by the SSG algorithm (right) [26]. The most homogenous reconstruction result can be achieved by method 1 (left), while method 2, especially for the SG algorithm, shows strong localized differences. The mean errors across the volume, after masking to remove noise and the corresponding histogram, are below each map, respectively.
Fig 6.
The g-factors of the different reconstruction approaches were calculated as suggested in [30] and displayed for completeness.
The mean g-factor across the object region are in the same range for the compared approaches, with the chosen reconstruction strategy (left) indicates a slightly higher g-factor than the two others. The corresponding histograms as well as the mean g-factor for all slices are displayed below each map.
Fig 7.
Synthetic SMS images (FA = 5°, TR = 30 ms) were disentangled with reconstruction kernels calculated from either ACS with identical (FA = 5°, TR = 30 ms) or different image contrast (FA = 90°, TR = 30 ms).
The resulting normalized signal differences for SG (left) and SSG (right) with respect to the base signal S0 are shown for three levels of SNR in the ACS (1, 5 and 8 averages). Subfigure (a) and (b) depict the first (TE1) and third echo (TE3) when reconstructed with a single, constant kernel (ACS111). In (c) each echo is reconstructed with its corresponding echo in the ACS (ACS123). The subfigures (d), (e) and (f) illustrate these dependencies for SSG. S0-values along the x-axis were binned to a width of 5. On the y-axis are the normalized differences between both reconstructions as formulated in Eq 6, the vertical bars represent the standard deviation of the mean value after binning.
Fig 8.
The T2*-dependency of the image signal for unwrapped, synthetic SMS images (FA = 5°, TR = 50 ms) is shown for SG (left) and SSG (right).
As in Fig 7 the reconstruction results of different ACS are compared by subtraction. While (a), (b), (d) and (e) represent reconstructions where ACS is from the first echo only (ACS111), the subfigures (c) and (f) show SG and SSG reconstructions for TE3 when each echo receives its own reconstruction kernel (ACS123). The bin-width is 5 ms. On the y-axis are the normalized differences between both reconstructions as formulated in Eq 6, the vertical bars represent the standard deviation of the mean value after binning.
Fig 9.
As shown in Figs 7 and 8, the robustness of the reconstruction with respect to S0 (red) and T2* (blue) was quantified by a linear model (b0, b1).
Besides different combinations of the ACS training data, three different SNR-levels of the ACS data are shown (1, 5 and 10 averages) for SG (left) and SSG (right). The slopes, b0, for both contrasts, S0 and T2*, are visualized in the top row for the compared training configurations and TEs. The bottom row displays the offsets, b1, of the linear model.