Quantitative T1 mapping using multi-slice multi-shot inversion recovery EPI

An efficient multi-slice inversion–recovery EPI (MS-IR-EPI) sequence for fast, high spatial resolution, quantitative T1 mapping is presented, using a segmented simultaneous multi-slice acquisition, combined with slice order shifting across multiple acquisitions. The segmented acquisition minimises the effective TE and readout duration compared to a single-shot EPI scheme, reducing geometric distortions to provide high quality T1 maps with a narrow point-spread function. The precision and repeatability of MS-IR-EPI T1 measurements are assessed using both T1-calibrated and T2-calibrated ISMRM/NIST phantom spheres at 3 and 7 T and compared with single slice IR and MP2RAGE methods. Magnetization transfer (MT) effects of the spectrally-selective fat-suppression (FS) pulses required for in vivo imaging are shown to shorten the measured in-vivo T1 values. We model the effect of these fat suppression pulses on T1 measurements and show that the model can remove their MT contribution from the measured T1, thus providing accurate T1 quantification. High spatial resolution T1 maps of the human brain generated with MS-IR-EPI at 7 T are compared with those generated with the widely implemented MP2RAGE sequence. Our MS-IR-EPI sequence provides high SNR per unit time and sharper T1 maps than MP2RAGE, demonstrating the potential for ultra-high resolution T1 mapping and the improved discrimination of functionally relevant cortical areas in the human brain.


Adiabatic inversion pulse: effect of efficiency on the accuracy of T1
We simulated the error measured on T1 for three target values (1200, 1600 and 2000 ms) for different efficiency levels of the adiabatic inversion pulse (Figure 2). The red line shows no error for a perfect inversion (efficiency=1, 180 o ) irrespectively of TR. When the inversion pulse is not perfect, T1 is overestimated but this error decreases with increasing TR: for efficiency=0.9 (corresponding to a flip angle of 162 o ) the error for a target T1 of 1200ms (~WM at 7T) decreases from 20ms for a TR=1s to less than 10ms for TR>3s. Although this error increases for longer T1 target values, the error for target T1 of 2000ms remains low (10.7ms) for the TR (5s) used in our experiments.

3.Measurements of ISMRM/NIST system T1 and T2-spheres.
Previous work using the ISMRM/NIST system T1-spheres had showed that the DREAM B1+ mapping sequence is only accurate for T1 values larger than approximately 300ms (Kato et al., 2019). Our B1-field measurements with the DREAM sequence inside each T1-sphere showed that this was also the case at 7T (Fig.3A). The B1-field measured inside spheres T1_7 to T1_14 at 7T was reduced from the nominal B1 by more than 50%, in spheres T1_5 and T1_6 the value of B1 was reduced by 67 and 58% respectively, whilst for spheres T1_1 to T1_4 the B1 was above 80% of its nominal value. One consideration for T1-mapping with the MP2RAGE sequence is that the MP2RAGE signal intensity is not monotonic for the chosen acquisition parameters at 7T (which was based on the sequence described by Haast et al. 2018). Since the MP2RAGE signal intensity values in the lookup table do not have a discrete one-to-one correspondence for low T1, some spheres may comprise voxels which are assigned T1 values on either side of the 540ms peak at 7T (Fig.3C), as it is the case for spheres T1_5 and T1_6 (~400 and 290ms respectively) where some voxels have been clearly misclassified, as shown by the noisy map within these spheres in Fig.3B. Histograms of MP2RAGE T1-values ( Fig.3D) show two peaks for T1_5 and T1_6, but also for T1_4 (this was also the case for T2 spheres T2_7, T2_8 and T2_9). Whenever two histogram peaks were present, a cut-off T1-value was applied to measure T1-values only across the voxels contributing to the correct peak (based on the knowledge of expected T1 reference value). Details of mean T1 values (across sessions) measured in NIST T1 spheres at 3T and 7T for each T1-mapping method, along with their respective NiCl2 concentration and reference 3T T1 values provided by NIST are provided in Table 1. Note a large variation of measured T1-value is apparent in the T1_3 sphere across time (but consistent values were measured with all T1-mapping methods for this sphere within a scan session); T1-values measured using the standard single-slice IR sequence were 966±17ms and 820±19ms at 3T and 7T respectively for session 1, and 1414±9ms and 1227±9ms at 3T and 7T respectively in the scan session 9 months later. Hence, for the T1_3 sphere, the accuracy was evaluated based on the first measurement alone, while the coefficient of variation (CoV) was evaluated on the last four measurements acquired within a 10 day window. Table 2 contains the corresponding information for the NIST MnCl2-doped T2 spheres. Note that the reference T1 values for spheres T2_1 and T2_5 are reported with respect to known deviations of the MnCl2 concentration due to an error in the manufacturing process.

Alternative method for correction of MT effects from the T1 quantification
An alternative method to compute the b/a parameter maps is to reduce the amount of fat suppression pulses, rather than varying the FA of the SPIR fat suppression pulse. On one subject we performed two acquisitions using SPIR fat suppression pulse with FA=90 o , one in which the SPIR pulse was applied prior to each RF acquisition (all) and one in which a SPIR pulse was applied prior to every alternate RF acquisition (alternate), hence using 50% Figure 4C shows original and corrected R1-maps for each method. The red arrowhead indicates a region of hyperintensity in the uncorrected R1-maps (due to spectrally selective FS pulse causing attenuation of the proton signal), which is more evident for FS pulses with higher flip angles. The corrected R1-maps however are homogeneous across the brain as the spatial variations are reflected (and corrected) by the fitted voxelwise b/a parameter map.

b/a parameter estimation results for subjects 4 and 5 (section 3.5)
Five subjects participate in the experiment to compare MS-IR-EPI with single slice IR-EPI in a single slice and MP2RAGE across the whole brain. Subject specific b/a parameter values were available for subjects 1, 2 and 3, who had participated in the experiment to model the impact of FS pulses (Section 2.2.1). For subjects 4 and 5, the b/a-parameter map was estimated from two data sets acquired with SPIR flip angles of 40 o and 70 o . Figure 5 plots histograms of b/a parameter values measured for subjects 3 and 4. The b/a parameter for the global correction was computed from the two Gaussian mixture fit, using the mode (peak) of the highest amplitude Gaussian (0.0043 and 0.0042 for subjects 3 and 4 respectively).

LCPA denoising of high spatial resolution data
Complex denoising was performed using a local complex principal component analysis (LCPCA) technique (Bazin et al., 2019) which has been shown to efficiently improve SNR of MR images acquired at the lower limit of SNR. Figure 6A shows the effect of de-noising for an example slice acquired at very high spatial resolution (0.35x0.35x0.7mm 3 ) with MS-IR-EPI (acquired at inversion time of 495ms) using the surface coil. The SNR improvement is very clear, particularly for the cerebellum where SNR was very low due to the positioning of the surface coil. The SNR improvement is not so dramatic for the derived T1-map (Figure 8 in main manuscript), probably due to the SNR improvement when combining the individual acquisitions (n=15) to compute the T1-map. Cortical profiles extracted from the T1 and M0 maps derived before and after de-noising are very similar (Fig. 6B).  (TR=3.2s,TE=20ms,EPI factor= 13,35 slices,offsets=[0,7,14,21,28], 3 averages, 11min 8s total acquisition time) for an example slice acquired with TI=495ms before (top) and after (bottom)