Yet more evidence that non-aqueous myelin lipids can be directly imaged with ultrashort echo time (UTE) MRI on a clinical 3T scanner: a lyophilized red blood cell membrane lipid study

Direct imaging of semi-solid lipids, such as myelin, is of great interest as a noninvasive biomarker of neurodegenerative diseases. Yet, the short T 2 relaxation times of semi-solid lipid protons hamper direct detection through conventional magnetic resonance imaging (MRI) pulse sequences. In this study, we examined whether a three-dimensional ultrashort echo time (3D UTE) sequence can directly acquire signals from membrane lipids. Membrane lipids from red blood cells (RBC) were collected from commercially available blood as a general model of the myelin lipid bilayer and subjected to D 2 O exchange and freeze-drying for complete water removal. Sufficiently high MR signals were detected with the 3D UTE sequence, which showed an ultrashort T 2 * of ~77 – 271 µ s and a short T 1 of ~189 ms for semi-solid RBC membrane lipids. These measurements can guide designing UTE-based sequences for direct in vivo imaging of membrane lipids.


Introduction
A significant portion of the molecular composition of cell membranes and myelin consists of the lipid bilayer.The typical cell membrane differs from myelin in the overall lower proportion of lipids and the ratio of three major classes of lipid components.The cell membrane has 25 % cholesterol, 65 % phospholipids, and 10 % glycolipids, while in myelin the ratio is closer to 40 %:50 %:20 %, respectively (Poitelon et al., 2020).Nevertheless, the lipid bilayer is the basic structure of all cell membranes and myelin.Myelin sheath is a modified cell membrane that wraps layer-by-layer around the nerve axon (Kister and Kister, 2022).The membrane lipids may serve as biomarkers of neurological diseases involving apoptosis and demyelination, such as multiple sclerosis (Devarajan, 2005;Lucchinetti et al., 2000).Currently, the integrity of membrane lipids can be probed only by biopsy, an invasive procedure prone to sampling errors.Thus, a noninvasive imaging method that can selectively monitor membrane lipids will be a valuable diagnostic tool for many neurological degenerative diseases.
Magnetic resonance imaging (MRI) has been widely studied for acquiring anatomical and morphological information and quantitating molecular composition and microstructural integrity via diverse contrast mechanisms.For monitoring membrane lipids and myelination levels in the brain, several MRI approaches have been suggested, such as magnetization transfer (MT) imaging, bound water (e.g., myelin water) imaging, and chemical exchange saturation transfer (CEST) imaging (Swanson et al., 2017;Lee et al., 2021;Zhao et al., 2023).Although these methods have demonstrated the potential of quantitatively measuring the membrane lipids and myelin, they are fundamentally indirect measurements that depend on the exchange between free water protons and membrane protons or on probing surrogate bound water pools associated with membrane lipids.This dependency on the exchange between two proton pools and surrogate markers complicates the quantification of membrane lipid and myelin content.Furthermore, it is difficult to evaluate the quality of membrane lipids and myelin, such as their T 1 and T 2 * relaxation times.
For direct imaging of membrane lipids, which possess very short T 2 relaxation times, ultrashort echo time (UTE) MRI pulse sequences can be used (Waldman et al., 2003;Horch et al., 2011;J. Du et al., 2014;Y.-J. Ma et al., 2020;Ma et al., 2022;Wilhelm et al., 2012).UTE MRI has been demonstrated as a useful platform for imaging short T 2 tissues such as bone, tendon, and myelin, as well as for characterizing their relaxation properties (Y.-J.Ma et al., 2020;Afsahi et al., 2022;J. Du et al., 2014;Y. Ma et al., 2020;Chu and Williams, 2019;Du et al., 2010).Myelin has been of great interest as a target for this technique to measure relaxation times in an effort to develop UTE imaging as a sensitive diagnostic tool for neurodegenerative diseases such as multiple sclerosis (van der Weijden et al., 2021).Various dehydrated ex vivo myelin samples were extensively studied to measure their T 1 and T 2 * relaxation times via these sequences.However, the measurements highly depend on the sample preparation and image acquisition/analysis methods (J.Du et al., 2014;Wilhelm et al., 2012;MacKay et al., 2006;E.L. Baadsvik et al., 2023).Myelin samples may also lose their unique multilamellar membrane structure during the preparation, which can significantly alter the MR relaxation properties (Shatil et al., 2018).Thus, to develop UTE as a robust and selective imaging method for myelin, the relaxation times should be measured with a more generalized form of membrane lipids that maintains the original structure during the sample preparation.
Here, we demonstrate the feasibility of directly imaging membrane lipids via the UTE sequence using red blood cell (RBC) membranes and measuring their relaxation properties.RBC membranes have been widely studied for preparing ghost membranes that preserve their original form (Schwoch and Passow, 1973;Deák et al., 2015).The prepared samples were completely dehydrated to ensure the measured MR signals were only from non-aqueous protons in the RBC membrane samples.T 1 and T 2 * relaxation times were measured using different analysis methods and compared with previous studies on myelin samples.

RBC membrane preparation
The RBC membranes were collected from commercially available blood (Lampire Biological Laboratories, Pipersville, PA).The plasma and buffy coat were removed by centrifugation, and the remaining RBCs were washed with Tris buffer (0.176 M, pH 7.6).Cell membranes were collected by mixing with Tris buffer (11.35 mM, pH 7.6) for 5 min at 4 • C and centrifuging at 20,000 g for 30 min.The collected RBC lipid membrane was suspended in D 2 O for 24 h, centrifuged, and freeze-dried overnight in a 1 mL syringe.The D 2 O suspension and freeze-drying were repeated twice to remove water completely, followed by air removal via compressing and capping.

MR image acquisition and analysis
The RBC membrane sample in the 1 mL syringe (length: 10.2 cm, inner diameter: 4.78 mm) was placed parallel to the scanner bore in a wrist coil and scanned at 3T (GE Healthcare Technologies Inc, Milwaukee, WI) at room temperature.Images were acquired using a 3D UTE sequence with TR = 30 ms, TE = 32 µs, acquisition matrix size = 64 × 64 × 20, field of view (FOV) = 80 mm, slice thickness = 10 mm, number of averages = 1, and sampling bandwidth = 125 kHz.The sequence was repeated with seven different echo times (TE = 32,64,150,200,400,600, and 800 μs) to measure the T 2 * of the RBC membrane lipid.T 2 * was measured via both exponential (mono and bicomponent) and bicomponent (e.g., long T 2 * and short T 2 * components) super-Lorentzian fitting of images acquired with 7 different TEs.Normalized root mean squared error (NRMSE) was used to evaluate the goodness-offit.
A variable flip angle (VFA) approach with the same UTE readout was used to measure the T 1 of the RBC membrane lipid, using nominal flip angles of 1 • , 5 • , 10 • , 15 • , 20 • , and 25 • TR and TE were fixed to 30 ms and 32 μs, respectively.Nominal flip angles used for T 1 measurement were corrected by considering T 2 * decay during radiofrequency (RF) pulse excitation (Carl et al., 2010).T 2 * measured from bi-exponential fitting was applied for this purpose, as the T 2 * decay model is based on the assumption of exponential decay.

Exponential fitting for T 2 * measurement
T 2 * relaxation time was first measured via conventional exponential fitting.A custom-written MATLAB (MathWorks, Natick, MA) code was used for the fitting.The same UTE dataset with 7 different TEs was fit twice using a mono-exponential model and then a bi-exponential model (long and short T 2 * compartments).Bi-exponential fitting was performed to examine whether residual free water remained in the sample.The trust-region algorithm was used for the fitting process, with the boundary for short T 2 * ranging from 0.1 µs to 10 ms.

Super-Lorentzian fitting for T 2 * measurement
As ultrashort T 2 * species such as myelin are known to have super-Lorentzian line shape, we also fitted the same UTE dataset with 7 different TEs into the super-Lorentzian model as previously described using a MATLAB code (Wilhelm et al., 2012;Weiger et al., 2020).Similar to bi-exponential fitting, super-Lorentzian fitting was also performed in a two-component model, with one component representing a residual free water pool with an exponential line shape.Similar to bi-exponential fitting, a trust-region algorithm was used for the fitting with the boundaries for T 2 * of super-Lorentzian pool ranging from 1 µs to 1 ms.

Data and code availability statement
The supporting data, the reported findings, and the code used to analyze the data are available from the corresponding author upon request.
The raw images for the T 1 measurement showed decreasing signal intensities with flip angles (Fig. 2A).Without flip angle correction, the T 1 was measured to be 127 ± 7 ms.Correcting flip angles using T 2 * measurement from bi-exponential fitting (T 2 * = 252 ± 51 μs) updated the T 1 value to 189 ± 7 ms (Fig. 2B).
Table 1 summarizes the measured T 2 * values of the RBC membrane lipids in this study and the T 2 * values of non-aqueous myelin lipid protons reported in the literature.The T 2 * values measured with the exponential and super-Lorentzian fitting are largely consistent with values reported in previous studies.The exponential model provides consistently longer T 2 * values than the super-Lorentzian model.

Discussion
In this study, we demonstrated the direct detection of membrane lipids and measurements of their MR relaxation times via UTE MRI on a clinical whole-body scanner.The detection of MR signals from completely dehydrated RBC membrane lipids parallels the previous UTE studies on bovine myelin lipid imaging (Horch et al., 2011;Wilhelm et al., 2012;Sheth et al., 2016).The ultrashort T 2 * of ~77-271 µs and short T 1 of ~189 ms for semi-solid RBC membrane lipids are also similar to the relaxation times for myelin lipids.Combining D 2 O-exchange with freeze-drying of RBC membrane lipid samples guarantees that the MR signals are only generated from semi-solid lipid protons.These results demonstrate that UTE MRI can image non-aqueous RBC membrane lipid protons directly.The relaxation measurements can guide the designing of UTE-based sequences for direct in vivo imaging of membrane lipids.
On top of demonstrating the feasibility of direct detection of semi-solid membrane lipid protons, we also measured the T 1 and T 2 * relaxation times of these protons via UTE MRI.The measurement of these relaxation times is crucial for optimizing the UTE sequence to achieve the highest possible signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) (Young et al., 2020).Considering the very short T 2 relaxation times of semi-solid protons, it is of critical importance to consider the following factors in the morphological and quantitative imaging of  membrane lipids.First, the echo time should be kept as short as possible to minimize short T 2 signal decay.Conventional spin echo and gradient echo sequences with TEs of several milliseconds or longer cannot detect much signal from semi-solid membrane lipid protons.Second, the data sampling window should be kept relatively short to minimize signal loss during spatial encoding (Rahmer et al., 2006).Radial ramp sampling used in UTE imaging typically suffers from more spatial blurring in imaging short T 2 tissues than zero echo time (ZTE) type sequences (E.L. Baadsvik et al., 2023;Weiger et al., 2020;Rahmer et al., 2006;Seifert et al., 2017).Third, the radiofrequency (RF) pulse should be kept short to minimize transverse relaxation during excitation (Carl et al., 2010).A high B 1 field is preferred for more accurate excitation of membrane lipids.A short rectangular pulse with the highest B 1 available on a clinical MR system is typically employed for direct imaging of myelin and membrane lipids (Wilhelm et al., 2012).Fourth, efficient water signal suppression is required for high contrast imaging of myelin and membrane lipids, as the water signal is far higher than that of myelin and membrane lipids (Fan et al., 2018).
The measured T 1 and T 2 * relaxation times can also be utilized to implement this UTE sequence for in vivo imaging of membrane lipids (Young et al., 2020).UTE imaging of short T 2 tissues in vivo often involves magnetization preparations such as inversion recovery or the post-processing of acquired images such as echo image subtractions to acquire positive contrast from the target short T 2 tissues (Jang et al., 2020;Ma et al., 2021).For the effective suppression of signals from long T 2 tissues, the adiabatic inversion recovery preparation followed by echo subtraction seems more efficient than other approaches, such as long T 2 signal saturation (Larson et al., 2006), or subtraction of longer echo images from short echo images (Weiger et al., 2020), largely because of the insensitivity of adiabatic inversion pulses to B 1 and B 0 inhomogeneities (Garwood and DelaBarre, 2001), thereby providing uniform inversion and nulling of water signals.However, the inversion recovery time and echo spacings should be carefully chosen based on the T 1 and T 2 * relaxation times of the tissues (Y.-J.Ma et al., 2020).The T 1 and T 2 * relaxation times of the RBC membrane from this study can serve as a good starting point for future studies on selective in vivo imaging of membrane lipids (Young et al., 2020).Still, it should be noted that the differences in membrane lipid compositions and microenvironment between different tissues (e.g., RBC membrane vs.myelin) can lead to different T 1 and T 2 * relaxation times, requiring fine measurements for each tissue type of interest.The magnetization exchange between different components can further complicate the measurement (especially T 1 relaxation time) and analysis of relaxation properties, as these components can be excited unequally by an RF pulse (Soustelle et al., 2023).We also measured the relaxation properties at room temperature.For in vivo imaging, it is reasonable to extrapolate that the T 1 will be longer than our measurements due to the higher temperature.The oxygen saturation level can be a confounding factor in vivo, as the reduction of T 2 * in the presence of deoxyhemoglobin is well known as the basis of the blood oxygen level-dependent (BOLD) effect.The iron in hemoglobin may also further shorten the T 2 * relaxation time.However, the blood still has a relatively long T 2 /T 2 * so that the signals from blood can be suppressed by using an adiabatic inversion pulse and an appropriate nulling time point, enabling selective UTE imaging of ultrashort T 2 * signals from RBC membranes (Ma et al., 2021; Y.-J.Ma et al., 2020).
We measured T 1 and T 2 * of RBC membrane lipids, which were around 189 ms and 77-271 μs, respectively.These measurements align with previous reports on the T 1 and T 2 * relaxation times of ex vivo myelin powder samples or brain white matter (Horch et al., 2011;Wilhelm et al., 2012;J. Du et al., 2014;MacKay et al., 2006;Weiger et al., 2020;Sheth et al., 2016;He et al., 2017;Sheth et al., 2017;Fan et al., 2017).This alignment of measurements was expected as both myelin sheath and RBC membranes share a similar structure of lipid bilayers.The discrepancy between T 2 * measurements from bi-exponential (252 μs) and bicomponent super-Lorentzian (77 μs) fitting was expected, as lipid bilayers show super-Lorentzian line shape and exponential fitting can overestimate the T 2 * values (Wennerström, 1973).Despite this difference in T 2 * measurements, both models show a similar high amplitude of short T 2 * components (97 % from bi-exponential and 95 % super-Lorentzian), cross-validating that the free water was almost completely removed via D 2 O exchange and freeze-drying.The T 2 * of 77 ± 5 μs from super-Lorentzian fitting falls within the range of previously reported myelin T 2 */T 2 values summarized in Table 1 (e.g., ~50 μs (MacKay et al., 2006), 7.5-101 μs (Weiger et al., 2020), 50-1000 μs (Horch et al., 2011), 150-250 μs (J.Du et al., 2014;Sheth et al., 2016;Y.-J. Ma et al., 2020;He et al., 2017;Sheth et al., 2017;Fan et al., 2017), or 8 μs-26 ms with ~90 % of the myelin T 2 * < 1000 μs (Wilhelm et al., 2012)).Yet, other studies used three-component super-Lorentzian fitting and revealed the existence of a very short T 2 * (~5 μs) component, which is unlikely to be detected by our UTE sequence with TE of 32 µs at a whole body scanner (E.L. Baadsvik et al., 2023;E.L. Baadsvik et al., 2023).This wide range of T 2 * values indicates the existence of multiple components in myelin and complicates the procedure of assigning the molecular origin of each T 2 * component.Similarly, RBC membrane components other than lipids, such as actin filaments, may also contribute to the UTE signals detected in this study, analogous to the fact that myelin basic proteins can also contribute to UTE signals of myelin.Different fitting approaches (mono vs. bi-component, exponential vs. super-Lorentzian) employed in this study may be useful for probing different components in RBC membranes and myelin based on their molecular mobility and T 2 * relaxation.
Considering that we used completely dehydrated RBC membranes, the solid-state NMR spectrum can help resolve the ambiguity of line shape, identify different components, and measure their T 2 * relaxation.
Meanwhile, a recent study by Seifert et al. demonstrated an excellent mono-exponential fit to the first 2 ms of 3T free induction decay (FID) data, yielding a T 2 * of 238 ms and R 2 of 0.994 31 .In contrast, the super-Lorentzian fit produced an R 2 value of 0.9992.The super-Lorentzian model only slightly increased the R 2 value (0.9992 vs. 0.994, or 0.5 % increase) over the mono-exponential fit at 3T.In other words, the muchsimplified mono-exponential model works very well in quantifying myelin T 2 * at 3T.It is likely that the super-Lorentzian model might be more accurate in fitting the supershort T 2 * (~10 µs) components, while the mono-exponential model might work well for the ultrashort T 2 * (~100-200 µs) components, as suggested by the equally well fitted  super-Lorentzian model and mono-exponential model in the Seifert study and our current study.The strong spatial blurring might be a key factor explaining the similar good fitting using both models.The sampling window is on the order of hundreds of microseconds in typical UTE or ZTE imaging, which is much longer than the supershort T 2 * components and significantly longer than the ultrashort components, leading to strong spatial blurring.A numerical simulation study might help explain this phenomenon, especially on whether the ~10 µs components could be spatially encoded and potential errors in T 2 * and T 1 quantification.
For the T 1 measurement, we used the VFA method with flip angle correction regarding the rapid T 2 * decay during the RF excitation.The elongated T 1 relaxation time after the correction indicates the need for the flip angle correction to measure T 1 relaxation time in the case of short T 2 tissues.The updated T 1 value of 189 ms roughly aligns with the previously reported T 1 relaxation times of 150-230 ms from the fourpool model analysis of white matter samples from a fresh bovine brain (Manning et al., 2021), and 197-248 ms from single exponential recovery fitting of long-T 2 suppressed UTE images of white matter in healthy volunteers (J.Du et al., 2014).To enhance the precision of VFA-based T 1 measurement, acquiring a reference amplitude and including higher flip angles can be helpful as we did not observe the peak amplitude in the VFA curve shown in Fig. 2. Other image acquisition and analysis methods can also be tested for cross-validation, such as the variable repetition time (VTR) method and multi-pool analysis (Ma et al., 2018;Reynolds et al., 2023).
We used the RBC membrane as a general model of membrane lipids in tissue.The protocol of preparing "ghost membranes" from RBCs using hypertonic solution has been well established over decades and is widely utilized in camouflaged nanoparticle synthesis these days (Schwoch and Passow, 1973;Luk and Zhang, 2015;Xia et al., 2019).This protocol is well suited for our study as the original structure of lipid bilayers of RBC membranes is relatively well preserved.The relaxation properties of myelin powder have been extensively studied in light of applications in neurodegenerative diseases such as multiple sclerosis.However, whether the unique multilamellar lipid bilayer structure of myelin is preserved in a powder form is uncertain, questioning whether the measurements from ex vivo samples are comparable to the true relaxation properties of semi-solid myelin protons in vivo.Although further studies are needed, the measurements from RBC membranes match well with previous reports on ex vivo myelin samples, implying that the loss of structural integrity of myelin did not heavily affect the previous T 1 and T 2 * measurements.
These relaxation measurements from RBC membrane samples can be used to optimize the UTE sequence for direct in vivo imaging of membrane lipids and myelin in general, owing to their structural and compositional similarities.Current MRI methods for detecting demyelination are either non-specific to myelin or indirect, probing general T 2 or T 2 * changes.Similar to approaches discussed above for general membrane lipid imaging, the same principles can be applied for direct imaging of myelin-suppressing long T 2 components utilizing the relaxation times measured in this study and directly detecting signals from myelin protons with UTE readout.UTE sequences have distinct advantages over conventional MRI sequences, which are indirect and have difficulty quantifying membrane lipid content, especially regarding T 1 and T 2 * relaxation times.This approach of directly imaging myelination may not only improve the diagnosis of neurodegenerative diseases but also facilitate the development of novel therapies for such diseases by offering reliable noninvasive methods of assessing therapeutic efficacy and safety profile.
In future studies, the accuracy of T 2 * measurement should be improved by acquiring more TEs.In this study, we only used 7 TEs for the T 2 * measurement.This T 2 * measurement will also affect the T 1 measurement through the VFA approach, as the flip angle should be corrected due to fast transverse relaxation during the RF excitation for ultrashort T 2 tissues.The T 1 was updated to 189 ± 7 ms from 127 ± 7 ms after the flip angle correction based on the T 2 * measurement via biexponential fitting.A signal decay model that accounts for super-Lorentzian line shape may allow more accurate flip angle correction and subsequent T 1 measurement.On top of acquiring more TEs, the chemical shift of membrane lipids should also be considered for more accurate T 2 * measurement.Each of the different chemical compositions and functional groups in the lipid membrane can be modeled as a separate super-Lorentzian function with different resonant frequencies.In this study, only one super-Lorentzian function was used for the fitting, and the relative chemical shift from the long T 2 component (modeled as an exponential function) was not considered for T 2 * fitting.Taking the chemical shift into account may reveal super-Lorentzian components with even shorter T 2 * relaxation times (Seifert et al., 2017).Furthermore, a Gaussian lineshape might be more appropriate for the supershort T 2 * (~5 µs) components.A systematic study should be performed to compare the Lorentzian, super-Lorentzian, and Gaussian lineshapes for different T 2 * components of the non-aqueous cell membrane and myelin lipid protons.
Lastly, the correction of UTE-associated spatial blurring (e.g., T 2 * and point-spread function (PSF) blurring) may lead to a higher accuracy in relaxation time measurements.PSF blurring is a well-known challenge in UTE MRI.In this study, no PSF correction was applied.All the UTE images with different flip angles and echo times were subject to similar PSF blurring.Future studies are needed to correct PSF blurring and assess its effect on relaxation time measurements.T 2 * blurring can be minimized via several methods, including high amplitude gradients, correction of phase and T 2 * along the readout, and T 2 -adapted sampling (Rahmer et al., 2006).

Conclusion
We demonstrated the feasibility of 3D UTE sequences to directly image the semi-solid membrane lipids.We also measured the T 1 (~189 ms) and T 2 * (77-271 μs) of the RBC membrane lipids that could be utilized to optimize UTE sequences for in vivo membrane lipid imaging.

Fig. 1 .
Fig. 1.Direct UTE imaging of red blood cell (RBC) membrane lipids and T 2 * measurements.(A) Raw images of the RBC membrane lipid sample at different echo times.A red region-of-interest (ROI) was used to measure T 2 * relaxation times via exponential fitting.(B) and bicomponent super-Lorentzian fitting (C).(T 2s * = T 2 * of short T 2 * component.a s = relative amplitude of short T 2 * component).

Fig. 2 .
Fig. 2. T 1 measurement of semi-solid RBC membrane protons via variable flip angle (VFA) method.(A) Raw images of RBC membranes over a range of flip angles were used for the T 1 measurement.(B) Comparison of T 1 measurement before and after correction for the T 2 * decay during RF excitation.The flip angle correction updates the T 1 measurement from 127 ± 7 ms to 189 ± 7 ms.

Table 1
Comparison of T 2 * relaxation time of red blood cell (RBC) membrane lipids measured in this study with previous studies on myelin lipids, ex vivo brain samples, and in vivo brain scans