Improved gray‐white matter contrast using magnetization prepared fast imaging with steady‐state free precession (MP‐FISP) brain imaging at 0.55 T

To improve the gray/white matter contrast of magnetization prepared rapid gradient echo (MP‐RAGE) MRI at 0.55 T by optimizing the acquisition and sequence kernel parameters.

generation of low-field 0.55 T scanners as well.In general, it is beneficial to reevaluate the design of a sequence when adapting it to different field strengths.When it comes to T 1 -weighted sequences, like the MP-RAGE, this is especially important at lower field strengths, due to the considerably shorter T 1 relaxation. 10,11he concept of an MP-RAGE sequence is simple 7 : A block of magnetization preparation, usually a 180 • inversion pulse, is followed by N rapid gradient echo (GRE) readout blocks, sampling k-space along the inversion curve.While the GRE kernel is typically a radiofrequency (RF) spoiled GRE (FLASH), in general, any type of readout might be utilized.A close sibling of the FLASH, and another possible candidate for this, is the FISP (fast imaging with steady-state precession). 12The main difference between the two is that, in contrast to a FLASH, there is no RF-spoiling in a FISP sequence.This means, the transverse magnetization is not destroyed at the end of each TR interval, leading to a different signal behavior. 12s mentioned, one of the main points to consider at lower field strength is the much shorter T 1 relaxation.The contrast in an MP-RAGE sequence is mainly controlled by the effective inversion time TI eff , which is the time between the center of the inversion pulse and the central k-space readout.It is limited by scan resolution but can be lengthened by introducing a delay before the readout (which is done at higher field strengts).For shorter T 1 values it is assumed that a shorter TI eff is beneficial for good contrast.However, an arbitrary shortening of TI eff can not be easily achieved due to the given number of readouts N (given by the resolution), which need to fit inside one readout block.A way to drastically reduce N, thereby shortening the inversion blocks, is to segment the sequence.This cuts N in half or even thirds, allowing for much shorter TI eff .
To test a segmented readout, as well as the FISP kernel, we first conducted a simulation of the signal contrast by employing the extended phase graph (EPG) formalism [13][14][15] to get some guidance for which parameters to choose.The T 2 -dependency of the two kernel types was compared.A custom sequence was then implemented which allows for the investigation of the aforementioned changes.It was compared to the product MP-RAGE with preset parameters in terms of signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR).Furthermore, a reproducibility study was carried out in the form of automatic tissue segmentation.
While the sequence design of an MP-RAGE with FISP readout technically still fits into the "RAGE" category, we renamed it to MP-FISP for easier discriminability.

EPG simulations
To have some guidance in finding the optimal parameters for maximum GM/WM contrast, a simulation was performed.For this, the MP-RAGE and MP-FISP signal were computed, using the EPG method, [13][14][15] as described by Kir et al. 16 The MP-RAGE signal of a tissue with a given proton density  and relaxation times T 1 and T 2 depends on the delay TD1 after the inversion pulse, before acquisition starts, the delay TD2 after acquisition, before the next inversion, the flip angle , the "small" repetition time  between  pulses (echo-spacing) and the number of  pulses N in one outer loop 16 (see Figure 1 for a visual depiction).
The EPG simulation keeps track of the state evolution by updating the expansion coefficients for each RF pulse i via matrix operations (see Kir et al., 16 equations [1] and [2]).Plugging in the tissue parameters (T 1 , T 2 , ) and sequence parameters (, TD1 and TD2, , N and RF phase Ψ i ) into these equations yields the signal evolution during data acquisition.RF spoiling is implemented according to the formula Ψ i − Ψ i−1 =  0 + i ⋅ , where  is the linear phase increment, leading to a quadratic phase increase with each RF pulse.The same computation can Basic diagram for an MP-RAGE pulse sequence introducing the used nomenclature.A magnetization preparation 180 • pulse is followed by a waiting period TD1 before data sampling starts.The time between excitation pulses with flip angle  is given by the "small" repetition time , each echo is acquired at TE after each pulse.N pulses are executed after each 180 • pulse.The effective inversion time TI eff is the time between the inversion pulse and the readout at k-space center i c .After the readout block there is a waiting time TD2 before the next magnetization preparation.The time for the whole outer loop, from 180 • pulse to 180 • pulse, is given by the "large" repetition time TR.
be used for both sequences, MP-RAGE and MP-FISP, by adjusting .While an MP-RAGE sequence is RF spoiled by using a nonzero , the MP-FISP does not employ RF spoiling,  = 0 • .For the MP-RAGE, a typical choice for  is 50 • , 16,17 which is used by Siemens 18 and, therefore, for our calculations as well.
N is typically determined by the resolution of the scan but can be strongly reduced by segmenting the sequence, which can halve N or even reduce it to a third.The number of RF pulses N per outer loop is then given by N = N tot ∕n seg where n seg is the number of segments.We chose to stick with the preconfigured resolution of the product MP-RAGE (see Section 2.4 for sequence parameters), so the total number N tot of partitions was fixed at N tot = 154.
is limited by how much time there is for sampling withing TR and by how many sampling points N need to fit into one TR.Due to adjustments to the custom sequence, we were able to increase  to 10.8 ms (see Section 2.3 for details) and, consequently, this value was used in the simulation as well.
TD1, TD2, n seg (and consequently N) and  were varied for the simulation as follows: • TD1 and TD2 each in the interval of (0, 2000) ms where a logarithmical spacing was used with a total of 50 values each • n seg = {1, 2, 3} (N = {154, 77, 51} accordingly) •  in the interval of (0, 60) • in linear steps of 2 The simulation was done for all combinations of a few values of T 1,GM = {700,750, 800,850, 900} ms and T 1,WM = {400,450, 500} ms.T 2,GM = 112 ms, T 2,WM = 89 ms, 11  GM = 1 and  WM = 0.92 16 were assumed.Because the sequence is T 1 -weighted, the used T 1 value is of much greater importance than the T 2 value, which is why a range of T 1 values was used, while T 2 was kept constant at the reported literature value.
The effective inversion time is given by TI eff = TD1 + N ⋅ i c , where i c is the ith pulse at which the central k-space line is acquired.The "big" repetition time between two inversion pulses is given by TR = TD1 + N ⋅  + TD2.
In order to penalize long, unnecessary scan times an additional term √ N ⋅ ∕TR was multiplied to the obtained signal.This ensures that the ratio of actual data acquisition to total scan time is as high as possible.
The contrast was calculated as 19 where S WM and S GM are the WM and GM signal, respectively.Because the contrast is dominated by data in the k-space center, the contrast evaluation of the simulation was done at the i c th pulse (at TI eff ).The parameters for the maximum value of Con were then retrieved and all parameters within 10% of the obtained optimum were examined.
The simulation was carried out with Matlab R2019a (The MathWorks, Inc.).

T 2 dependency
The MP-RAGE was designed in an effort to strive for superior, pure T 1 -weighting.This is ensured by the inversion pulse and by the choice of a FLASH readout as a typical T 1 -weighted sequence.It needs to be noted that pure weighting can not be achieved in any realistic case.Even a T 1 FLASH has some T 2 (or rather T * 2 ) contributions due to the finite echo time and choice of .However, the question arises, weather the undeniably high degree of T 1 -weighting of the MP-RAGE is retained when swapping the FLASH for a FISP, where the residual transverse signal is not destroyed by the following RF pulse and carries over to the next TR interval.In order to assess the amount of T 2 -weighting introduced by a FISP readout in comparison to a FLASH, a simulation was carried out.For this, the same EPG simulation as described above was used.In this case, T 1 ,  and all sequence parameters were kept constant, while T 2 was varied.Used tissue parameters were T 1,GM = 700 ms, T 1,WM = 450 ms,  GM = 1 and  WM = 0.92.The constant parameters were matched to the used sequence parameters:  = 20 • , TD1 = 0 ms, TD2 = 0 ms,  = 10.8 ms.An unsegmented (n seg = 1, N = 160) and a segmented (n seg = 2, N = 80) MP-RAGE and MP-FISP were computed.

Custom sequence
As described in the introduction, a custom sequence was implemented at the scanner.It is based on an ultra-fast SSFP sequence 20 and has the following relevant features: 1 Optimized gradient switching and slew rates for maximum readout time within . 2 Free choice of kernel type (FLASH or FISP): For testing how a different readout type performs at lower field strength.3 Possibility to segment the sequence: This provides the opportunity to chose a smaller N, even for higher resolutions.A smaller N, and therefore a shorter TI eff , could be beneficial for the contrast at 0.55 T due to the faster T 1 relaxation of tissue at this field strength.
4 Option to adapt the (inner-outer) loop encoding structure: For a 3D MP-RAGE, there are two possibilities when it comes to phase encoding: either the slice (partition) encoding is executed by the N inner loops and the in-plane phase encoding (lines) is done by the outer loop or the other way around.Depending on this, N is either the number of partitions or the number of lines.Usually, the former is used in product sequences, 12 but swapping of encoding directions can be used to slightly adjust N. We implemented the option to choose this freely in our custom sequence, as it slightly increases or decreases TI eff , leading to potentially better contrast.Note that, in theory, this could be achieved by swapping the encoding directions during acquisition.However, this would result in the need to adjust other geometry options, as well.
Especially for a direct comparison of sequences, this is disadvantageous.Note also, that since the influence on the value of N is rather small and the simulation is just meant as a guideline, this was not considered in the simulation but rather only tested experimentally at the scanner.
The overall goal was to use the product MP-RAGE as a starting point and explore if there is room for improvement.Therefore, the custom sequence was adjusted to the settings of the product MP-RAGE in terms of resolution, partial Fourier factor and filters, while making use of the above features.
The oversampling which is used in the product sequence was abandoned in the custom sequence, while the total scan time was kept roughly the same as the product sequence.This leaves TR unchanged, while N is reduced.Since less sampling steps need to fit into TR, this yields more time for each individual encoding step and allows for a longer .This is expected to be beneficial at lower field strength since there is less signal loss over  due to the longer T * 2 and the signal can be sampled more efficiently.Furthermore, the small delays TD1 and TD2 were reduced to zero (see Section 3 for in-depth justification) and a nonselective pulse was used instead of a slab-selective one.This allows to further increase the time within TR, which can be used for sampling.Due to these adaptions, we were able to increase  from 9.92 to 10.8 ms and lower the bandwidth of the custom sequence to 93 Hz/Px.The product sequence uses a bandwidth of 130 Hz/Px which coincides with the lower hard limit set by the vendor.The theoretical increase of SNR (and CNR) from lowering the bandwidth is simply obtained by calculating √ bw prod ∕bw cust , where bw prod and bw cust are the bandwidth of the product MP-RAGE and the custom MP-RAGE, respectively.Plugging in the used values yields a theoretical improvement of 18%.

Experimental comparison of sequences
To verify the results from the simulation, in vivo brain scans were performed on healthy volunteers.All imaging was performed at 0.55 T (MAGNETOM Free.Max, Siemens Healthineers).Scanning was approved by the local ethics committee and written informed consent was given by all volunteers beforehand.
Three sequences were compared in detail: • product MP-RAGE with preset parameters, TI eff = 858 ms (MP-RAGE prod 858) • custom MP-RAGE with the same parameters as the product, but with reduced bandwidth, par-in-lin, TI eff = 832 ms (MP-RAGE* nseg 832) • custom MP-FISP with the same parameters as the custom MP-RAGE, but with FISP readout, n seg = 2, lin-in-par, TI eff = 572 ms (MP-FISP* 2seg 572) The terms in brackets represent the short names which were used later in Table 3 for better discrimination.The asterisk marks all custom sequences.Since TI eff is the most important and intuitive parameter for tissue contrast, it was chosen as the name-giving feature.Table 1 shows all important parameters for the three sequences.They were tested on five volunteers.
To get a broader overview and some confirmation of less relevant simulation results one of the volunteers was scanned with the following sequences in addition to the ones above: The obtained data from this scan is not presented in detail and not shown visually, but it is used to substantiate simulation results.It is termed "test scan" in the following.Note that the simulation was also done for n seg = 3 but we excluded this setting from the test scan due to its unpromising results from the simulation.For all scans, the scaling factor on the custom sequence was set such that all sequences had the same noise level.

T A B L E 1
Relevant parameters for the three important sequences: product MP-RAGE (MP-RAGE), custom, bandwidth-optimized MP-RAGE (MP-RAGE*) and custom MP-FISP with n seg = 2 (MP-FISP*).Note: Values marked with a hyphen have the same value for all three sequences and are given in the column of the product MP-RAGE.

Analysis of experimental comparison data
All data were reconstructed at the scanner and analysis was done on the reconstructed data.The SNR and CNR of all data were calculated and compared.For this, the reconstructed data was fed into FastSurfer 21 for automatic segmentation of cortical and subcortical structures.FastSurfer replicates anatomical segmentation of the widely used open-source software package FreeSurfer, 22 by employing deep learning for much faster results.
For the SNR, the average WM value over the whole brain was divided by the average value  avg of the SD of noise in the air of the four corners of the image.For the CNR, four regions of interest (ROIs) were picked from FastSurfer: Cerebellum, putamen, caudate nucleus and superior frontal gyrus.The CNR for these regions was calculated as where S WM∕GM is the signal strength in the WM and GM region, respectively.For the cerebellum the CNR was calculated from the GM and WM inside the cerebellum, for the other three GM structures, the CNR was calculated from the GM value inside these structures and the average WM value of the whole brain.Since all data was sampled and reconstructed identically, this direct SNR and CNR comparison is appropriate.

Reproducibility of automatic segmentation
In order to assess how the proposed MP-FISP* sequence performs in an application, compared to the product MP-RAGE, a reproducibility study was carried out.Three volunteers were scanned five times with both sequences.They were taken out of the scanner in between each scan and were repositioned anew.The data was fed to fsl 23 FAST 24 for automatic WM, GM, and cerebrospinal fluid (CSF) segmentation.The resulting tissue volumes were computed with the fslstats function within fsl.To obtain a quantitative measure, the SD of the mean of each volunteer's volumes was calculated and divided by the mean, in order to assess the relative scatter between the datasets.
Note that we chose to not use the data from the Fast-Surfer segmentation.While FastSurfer (as well as the subcortical segmentation of fsl [FLIRT 25 ]) employs a registration to standard space and an atlas to segment the brain images, fsl FAST relies on grayscale values, 24 posing a better measure for the quality of the CNR.

EPG simulations and test scan
A general result of all simulations was that small values of TD1 and TD2 were favorable.In most cases, the simulation yielded zero as the optimal value for both, TD1 and TD2 and there was almost no variation in signal and contrast for small values of TD1 and TD2.
The obtained value for the flip angle was  = 20 • for almost all settings.Some combinations of T 1,GM and T 1,WM yielded slightly lower or higher flip angles as the optimum but these were all in the range of  = (18, 22) • and there was not much contrast variation in this range (≈ 2%). = 20 • was also the pre-set flip angle for the product MP-RAGE and it was decided to use this for all in vivo measurements.
In general, the variation of T 1 in the used range did not have a large influence on the results.This shows that the optimized parameters are relatively robust against small changes in tissue parameters, which is a valuable advantage.
Table 2 shows the SNR and CNR values for the test scan with all nine sequences, which are referenced at some points in the following paragraphs.The given values are average values from the four regions shown in Figure 2 (cerebellum, caudate nucleus, superior frontal gyrus, and putamen).

MP-RAGE
The obtained parameters from the MP-RAGE simulation came very close to the preset scan parameters of the product MP-RAGE on the scanner.The preset delays were set Notes: All values are averages over the four contrast regions, shown in Figure 2. The given uncertainty is the SD of the mean.
(A) (B) Example slices for the automatically segmented regions of interest (ROIs) by FastSurfer.The chosen regions were: cerebellum (Ce) (white matter [WM] and gray matter [GM] ROI is shown), caudate nucleus (CN) (both drawn in (A)) superior frontal gyrus (SFG) and putamen (Pu) (both drawn in (B)).The WM whole brain mask overlay is not shown to avoid clutter.The according averaged SNR and CNR values for the test scan can be found in Table 2, the detailed CNR for all regions and volunteers of the scans with the three compared sequences can be found in Table 3 as well as the averaged CNR and SNR.
to very low values of TD1 = 14 ms and TD2 = 10 ms (this is indirectly visible from TR and TI eff , which are the adjustable parameters in the product MP-RAGE) and the flip angle was already set to the obtained optimal angle of  = 20 • from the simulation, as mentioned above.
The returned optimal value for N was 154 (no segmentation).n seg = 2 yielded lower contrast but was within 8% of the optimal contrast.This means, segmenting the MP-RAGE does not have a very large influence on the contrast according to the simulation.This was validated by the test scan, which showed no significant difference in CNR and SNR between the two sequences.The change of TI eff from 832 ms (par-in-lin) to 1145 ms (lin-in-par) had also no noticeable influence on both, SNR and CNR, in the test scan.

MP-FISP
While the contrast was not predicted to be better when comparing the simulation result of the nonsegmented MP-FISP to the equivalent MP-RAGE (same bandwidth, flip angle, TD1 and TD2), the computed signal strength was on average 26% higher, according to the simulation.This trend was verified by the test scan, although it was not as high as in the simulation.The nonsegmented MP-FISP* had 9% (TI eff = 832, par-in-lin) and 19% (TI eff = 1145, lin-in-par) more SNR than the non-segmented MP-RAGE*, while the CNR decreased slightly.Contrast-wise, the MP-FISP did benefit from a segmentation.An average contrast increase of 24% for the MP-FISP with n seg = 2 over the unsegmented MP-FISP was obtained, while a segmentation with n seg = 3 showed again a decrease in contrast by approximately 15% compared to n seg = 2.The test scan verified this trend, yielding a 40% CNR increase for the segmented MP-FISP* with TI eff = 410 ms over the nonsegmented one with TI eff = 832 ms (both par-in-lin) and an even higher increase of 74% when comparing the lin-in-par MP-FISP*s (TI eff = 572 ms, segmented versus TI eff = 1145 ms, unsegmented).
However, the signal strength showed a significant decrease for increasing n seg .For both, the simulation and experimental data, this amounted to 29% on average for n seg = 2, compared to the nonsegmented.Coincidentally, the nonsegmented MP-RAGE* and the segmented MP-FISP* with n seg = 2 exhibited the same signal strength in the simulation and almost the same signal strength in the test scan when considering the TI eff = 572 ms (lin-in-par) MP-FISP* (the TI eff = 410 ms MP-FISP* had lower signal strength).This means, the segmented lin-in-par MP-FISP* with TI eff = 572 ms yields a better contrast than the MP-RAGE*, while retaining almost the same signal strength.We conclude that the optimal TI eff lays somewhere in the region of 572 ms.
Consequently, an in-depth comparison of the product MP-RAGE, the custom nonsegmented MP-RAGE* and the lin-in-par segmented MP-FISP* was performed.

T 2 dependency
Figure 3 shows plots for the T 2 dependency simulation.
The MP-RAGE and MP-FISP signal is drawn for GM and WM for the two different parameter settings.(A) and (B) display the GM and WM signal, respectively, for the unsegmented case, (C) and (D) display the ones for the segmented sequences.It is observed that both signals depend on T 2 in all cases.However, the MP-FISP does exhibit slightly more variation than the MP-RAGE in (B) and (C).A difference is observed between the unsegmented and segmented plots, where the segmented ones exhibit a stronger T 2 -dependence for both, the MP-RAGE and MP-FISP.Apparently, using a FISP readout at 0.55 T does not per-se introduce significantly more T 2 -weighting than the default FLASH readout already exhibits.The more prominent factor seems to be the segmentation.However, the variation is generally very low (maximum 6% over a T 2 variation of 30 ms) and does not introduce a significant amount of T 2 -weighting.

In-depth comparison of MP-RAGE, MP-RAGE*, and segmented MP-FISP*
For the reasons presented in the previous section, the segmented (n seg = 2) MP-FISP* with TI eff = 572 ms was chosen as the most promising sequence to be compared with the product MP-RAGE and the custom standard MP-RAGE*.

F I G U R E 4
Comparison of the three sequences (product MP-RAGE (A), custom MP-RAGE* with lower bandwidth (B) and segmented MP-FISP* (C)) in axial, sagittal and coronal orientation.The scaling factor was chosen such that the noise level of all three scans is the same and all scans are windowed equally.The SNR improvement from product MP-RAGE to custom MP-RAGE* is evident.A visual improvement in gray/white matter contrast can clearly be observed for the segmented MP-FISP* over the MP-RAGE.The different CNR values from the regions shown in Figure 2 can be found in Table 3 for all scans in which the three sequences were compared.For all regions, an increase in CNR is observed for MP-RAGE* over MP-RAGE and again MP-FISP* over MP-RAGE*.The average CNR increase varies between regions.The lowest total average improvement was observed in the putamen (31%), the highest in the cerebellum (58%).Furthermore, the average CNR over all regions and all volunteers is displayed as well as the averaged improvement between the sequences.An average CNR improvement of approximately 15% was observed when comparing the MP-RAGE* to the product MP-RAGE.Further CNR improvement for the segmented MP-FISP* over the custom MP-RAGE was 22%, which matches the expectation from the simulation.Note, however, that the simulation compared par-in-lin MP-RAGE and MP-FISP, while the experimental MP-FISP* was measured lin-in-par.This should not make a big difference though.
The average SNR for all three sequences is also given in Table 3.As expected, the SNR of MP-RAGE* is much better than for the product MP-RAGE (25% increase).This even surpasses the predicted 18% from the calculation.While the SNR does decrease a bit (−9%) for the segmented MP-FISP* when compared to the MP-RAGE*, the CNR improves by 22%.In total, a 41% improvement in CNR and a 14% improvement in SNR for the segmented MP-FISP* over the product MP-RAGE was obtained in this experiment.

T A B L E 4
Results from the reproducibility study.

Reproducibility comparison
Table 4 shows the results of the reproducibility comparison of the product MP-RAGE and the MP-FISP*.The mean volume, relative error, and relative error difference between the two sequences are presented.All obtained volumes for both sequences lie in the range of previously reported values from MRI studies. 26The spread of values within the same volunteer is not very high in general, all relative errors are below 2%, often even below 1%.There is no clear evidence of one sequence performing better than the other when it comes to automatic segmentation.The highest relative error differences (up to 0.7%) are for CSF where a variance between data sets can also be attributed to slightly different brain extractions instead of contrast issues.

DISCUSSION
In this work we have explored the potential improvement in GM/WM contrast of the MP-RAGE at 0.55 T. The above suggested changes of optimizing parameters and changing the sequence kernel to a FISP readout type lead to a substantial average improvement of 41% in CNR.In the following we discuss some points to be noted.

Simulation
As previously reported by other groups which did MP-RAGE simulations (e.g., References 16 and 19), the optima regions for the parameters are quite flat and leave some freedom for parameter choices.In general, a simulation can only serve as support due to a great variety of physiological effects such as partial volume, magnetization transfer, etc.Therefore, we refrained from performing a rigor, in-depth simulation and relied more on the obtained experimental data, while using the simulation as guideline.

Significance of T 2 dependency
While our simulations have shown that more T 2 -weighting is introduced by using a segmented MP-FISP, instead of an unsegmented MP-RAGE, it is unclear if this small increase is significantly disadvantageous.Visually, all obtained images looked like typical T 1 -weighting.Further studies on pathological images would be beneficial to determine if the slightly increased T 2 -weighting has negative impact on diagnosis.Since the simulations have shown a very similar T 2 dependency of MP-RAGE and MP-FISP in the unsegmented case, we would like to mention the possibility of using an unsegmented MP-FISP with similar CNR values as the product MP-RAGE, but with much higher SNR (see Table 2), in the unlikely case that the T 2 -dependency does prove to be highly disadvantageous.

CNR improvement
There is some variation in the improvement of CNR between regions and also volunteers.From Table 3 it is evident that some regions profit more than others.While the average improvement in the putamen was only 31%, the CNR in the cerebellum improved by 58%.For the same region, values do vary somewhat between volunteers, as well (see e.g., putamen, V3: 42% vs. V5: 22% improvement).However, even for the worst values (CH and putamen, V5) a total improvement of 22% was obtained.Therefore, we can confidently say that using a segmented MP-FISP definitely has great potential in increasing CNR for the MP-RAGE at 0.55 T.

Segmentation
Segmenting the readout does not only provide a better CNR, it has also the potential to reduce artifacts for motion-sensitive GRE sequences, such as FISP. 27Due to the long T 2 , especially CSF is one of the major sources of motion-related artifacts, such as ghostings, in the steady state of FISP.For MP-FISP two major factors lead to a strong mitigation of its prominent motion sensitivity.Firstly, the inversion imaging is performed in the transient rather than in the steady state.Secondly, the frequent inversion pulses lead to the saturation of long T2 components, like CSF.Shortening TR by segmentation helps to confine imaging to the transient state.In contrast, for long TR, that is without segmentation, the steady state might be reached during inversion recovery.Additionally, the frequency of inversion is doubled (with n = 2), leading to stronger saturation effects of CSF.For MP-FISP, CSF becomes almost entirely saturated appearing strongly hypointense, in contrast to its steady state that would yield a strongly hyperintense signal. 12Consequently, the typical motion sensitivity as commonly associated with FISP can be successfully mitigated using segmentation.Furthermore, it is worth mentioning that segmenting the sequence without highly increasing the scan time is, of course, only possible if the delay times TD1 and TD2 are set to zero.Otherwise, it makes sense to put as many RF pulses as possible into one outer loop to reduce the number of outer loops and, therefore, the pure waiting time without data acquisition.At higher field strengths, where nonzero delays are advantageous for signal strength and contrast, 16,19 a segmentation would take a great toll on the scan time in addition to the fact that it might not even be beneficial to the contrast in the first place.

B 1 homogeneity
While it was not part of this study and has not been discussed so far, it is noteworthy that an enormous advantage of low-field MRI lays in the field homogeneity.A big problem of MP-RAGE imaging at high-field is the inhomogeneity of B 1 .This lead to the invention of the MP2RAGE 28 which utilizes two readout blocks at different TI eff and  in order to negate this effect.An MP2RAGE acquisition takes almost double the time of an MP-RAGE.At lower field strengths, the B 1 field is much more homogeneous and the image homogeneity does not suffer.Therefore, there is no need for the MP2RAGE and an optimized MP-RAGE is of even greater value.

Benefit of MP-FISP
While the reproducibility comparison did not produce clear evidence that the obtained 41% are necessary for automatic tissue segmentation, we do believe that some people would appreciate a slightly clearer view when it comes to low-field T 1 -weighted brain images.It should also be noted that this study was carried out on healthy volunteers, where automatic segmentation is not particularly challenging.While a patient study was beyond the scope of this study, it could lead to a more conclusive result to investigate both sequences on, for example, lesion segmentation.In any case, we note that the proposed implementations are easy to carry out and lead to a clearly visible improvement of image quality while no major disadvantages could be observed in the present study.It is also worth noting again that, according to Table 2, there is the possibility of using a nonsegmented MP-FISP* with similar CNR than the product MP-RAGE, but with significantly more SNR.If the increased contrast is not desired, the usage of the custom unsegmented MP-FISP* can yield 38% more SNR for the same CNR as the product MP-RAGE.

CONCLUSION
We have shown that there is room for improvement when it comes to the MP-RAGE GM/WM contrast at lower field strengths.An optimized sequence sampling and the switch to a FISP readout have been shown to have clear potential in comparison to the product MP-RAGE.Looking at the current trend toward low-field scanners and their intrinsically lower signal strength, it does make sense to investigate and implement such small and easy modifications to ensure high-quality scanning.

Figure 4
Figure 4 shows a visual comparison of the three sequences ((A) product MP-RAGE, (B) custom MP-RAGE*, (C) segmented MP-FISP*) in all three directions: axial, sagittal, and coronal.The images are all windowed equally.A visual improvement can clearly be observed.Especially the signal intensity improvement from the bandwidth adjustment and an improvement in contrast towards the segmented MP-FISP* are visible.The different CNR values from the regions shown in Figure2can be found in Table3for all scans in which the three sequences were compared.For all regions, an increase in CNR is observed for MP-RAGE* over MP-RAGE and again MP-FISP* over MP-RAGE*.The average CNR increase varies between regions.The lowest total average improvement was observed in the putamen (31%), the highest in the cerebellum (58%).Furthermore, the average CNR over all regions and all volunteers is displayed as well as the averaged improvement between

5 mm 3 ) Rel. error (%)
Notes:The obtained mean volumes, as obtained by fsl FAST, for cerebrospinal fluid (CSF), gray matter (GM), and white matter (WM) are reported for all three volunteers.The given uncertainty is the standard deviation of the mean.The relative error (rel.error) was calculated by dividing the uncertainty by the mean.The relative error difference (rel.error diff.) of the two sequences was obtained by subtracting the relative error of the MP-FISP* from the one of the MP-RAGE.This means positive values show that the MP-FISP* performed better while negative values indicate that the MP-RAGE performed better.