Sensitivity of unconstrained quantitative magnetization transfer MRI to amyloid burden in preclinical Alzheimer’s disease

Conformational abnormalities in the amyloid β (Aβ) protein are one of the defining pathological hallmarks of Alzheimer’s disease (AD) (Thal et al., 2002), along with tau abnormalities and neurodegeneration (Jack et al., 2018). In the extracellular space, 40–42 amino acid Aβ peptide fragments cleaved from amyloid precursor protein, known as Aβ₄₀ and Aβ₄₂, form soluble aggregates known as oligomers, which further aggregate into insoluble fibrils and ultimately plaques (Karran et al., 2011). It is believed that several species of Aβ aggregates have neurotoxic or inflammatory effects (Heppner et al., 2015) that ultimately lead to tau hyperphosphorylation (Oddo et al., 2006), neurodegeneration, and cognitive symptoms (Bennett et al., 2004), though the exact mechanisms are unclear. Aβ accumulation begins in the neocortex (Fantoni et al., 2020), even in clinically asymptomatic participants, and can occur decades prior to the possible symptomatic onset of Alzheimer’s dementia (Jack et al., 2010). Identifying individuals with a substantial cortical amyloid load is necessary for many clinical reasons, such as the specific diagnosis of AD versus other dementias (Jack et al., 2018), as Aβ deposits may be found incidentally in other forms of dementia but are not their primary pathological feature. Additionally, quantification of amyloid burden enables the study of the impact of new disease-modifying therapies—such as emerging antiamyloid immunotherapeutics—which are being used to treat early stages of AD and increasingly being studied in preclinical patient populations (Yadollahikhales & Rojas, 2023).

The gold standard for noninvasive in vivo amyloid assessment is positron emission tomography (PET) (Chapleau et al., 2022; Johnson et al., 2013; Rabinovici et al., 2023). The second generation of amyloid PET tracers includes [¹⁸F]florbetaben, which has been demonstrated to have high sensitivity for fibrillar amyloid (Syed & Deeks, 2015). While cerebrospinal fluid (CSF) Aβ biomarkers are similarly sensitive and lumbar punctures enable simultaneous access to tau biomarkers without requiring an additional procedure (Huszár et al., 2024; Palmqvist et al., 2015), intravenous radiotracer injection followed by PET imaging is less invasive. Additionally, PET offers spatial localization of the Aβ signal which may confer the ability to detect regional amyloid depositions before pathological changes occur in the global neocortical signal. However, PET has several drawbacks, including cost, the need for specialized equipment, availability, limited spatial resolution, off-target binding, and ionizing radiation exposure.

An alternative, magnetic resonance imaging (MRI)-based technique that may be directly sensitive to the accumulation of extracellular protein deposits, such as Aβ aggregates, is known as quantitative magnetization transfer (qMT) (Henkelman et al., 1993). MT methods sensitize the MRI signal to a “semisolid” spin pool consisting of protons bound in large macromolecules such as lipids (e.g., myelin) and proteins (e.g., both Aβ₄₀ and Aβ₄₂ aggregates), which exchange with protons bound in the usual “free water” pool. MT’s sensitivity to the accumulation of insoluble Aβ plaques (note that soluble Aβ species contribute minimally to the MT effect) was previously demonstrated in transgenic mice (Bigot et al., 2014; Pérez-Torres et al., 2014; Praet et al., 2016), and in vivo studies using quantitative MT—which disentangles the biophysical contributions to the MT contrast—suggest that the “forward magnetization exchange rate” is the most discriminatory qMT biomarker for classifying AD versus controls and the conversion from amnestic mild cognitive impairment (MCI) to AD (Duan et al., 2022; Giulietti et al., 2012; Makovac et al., 2018).

It is unclear from prior qMT studies whether the observed differences in the “forward exchange rate” arise from changes in the exchange rate (e.g., due to the insolubility of Aβ plaques) or the macromolecular pool size (e.g., from neurodegeneration). Additionally, prior studies usually constrain the value of the difficult-to-estimate semisolid pool’s longitudinal relaxation rate $R1s$⁠. In this work, we quantify all parameters of the unconstrained 2-pool qMT model using a tailored hybrid-state pulse sequence (Assländer, Mao, et al., 2024). Furthermore, due to previous studies’ focus on clinically diagnosed MCI or AD (Duan et al., 2022; Giulietti et al., 2012; Makovac et al., 2018), it is unclear whether preclinical AD pathology can be detected with qMT biomarkers. This distinction is important because Aβ has been shown to accumulate well before the clinical manifestations of dementia (Jack et al., 2010). Therefore, we focus on the preclinical population in this study. Our central hypothesis is that amyloid accumulation in the preclinical population is detectable by qMT biomarkers—including $R1s$—due to the distinct biochemical properties of Aβ plaques.

We use the unconstrained qMT model (Fig. 1) presented in Assländer, Mao, et al. (2024), which describes the Bloch–McConnell equations (McConnell, 1958) of the 2-pool spin system (Henkelman et al., 1993):

where $ωz$ is the off-resonance frequency and $ωy$ is the Rabi frequency of the radiofrequency (RF) pulse which has a flip angle $α$ and duration $TRF$⁠. The magnetization of protons bound in free water (superscript ^f) is described in Cartesian coordinates by $xf,yf,zf$⁠. The free pool’s magnetization exchanges at the rate $Rx$ with the semisolid spin pool (superscript ^s), whose magnetization is captured in Cartesian coordinates by $xs,zs$⁠. Each pool has its own fractional size (⁠$m0f,s$⁠, where $m0f+m0s=1$⁠), longitudinal (⁠$R1f,s$⁠) and transverse (⁠$R2f,s$⁠) relaxation rates (inverse of times $T1,2$⁠), amounting to six core parameters: $m0s,R1f,R2f,Rx,R1s$⁠, and $T2s$⁠. The nonexponential decay characteristics of the semisolid pool in the brain’s parenchymal tissue are described by the super-Lorentzian lineshape (Morrison et al., 1995), where we use the time $T2s$ (instead of the rate) for consistency with the qMT literature. These decay characteristics are incorporated in the Bloch–McConnell equation with the generalized Bloch model (Assländer et al., 2021). For computational efficiency, we approximate the nonexponential decay characteristics with an effective exponential decay using the linearized relaxation rate $R2s,l(R2s,α,TRF)$ that results in the same magnetization at the end of an RF pulse with the flip angle $α$ and pulse duration $TRF$⁠. Without loss of generality, we neglect $ys$ assuming $ωx=0$ and $R2s,l≫ωz$⁠. More details regarding the generalized Bloch model are described in Assländer et al. (2021). Note the slight differences with respect to quantities often referenced in the MT literature, including the pool-size ratio $PSR=m0sm0f$ (Henkelman et al., 1993), where $m0s=PSR1+PSR$⁠, and the “forward exchange rate,” which is the product $m0s⋅Rx$⁠. For the latter, many previous studies could not estimate the parameters $m0s$ and $Rx$ separately due to limitations of the employed encoding strategies and signal models (Ramani et al., 2002), which is overcome here by the use of a tailored hybrid-state pulse sequence.

The highly restricted motion of macromolecules leads to an ultrashort $T2s≈10μs$⁠, which prevents direct detection of this pool with the typical echo times achievable on clinical MRI scanners (i.e., without hardware modifications). Hence, the semisolid pool can be detected on clinical MRI scanners only indirectly via its exchange with the free pool, impeding the estimation of $R1s$⁠. Consequently, authors have typically assumed $R1s​=1$ s (Giulietti et al., 2012; Henkelman et al., 1993; Makovac et al., 2018). In this work, we overcome these limitations by utilizing a hybrid-state sequence (Assländer, Mao, et al., 2024; Assländer et al., 2019) (more details given in Section 2.4). We recently demonstrated in vivo the ability to voxel-wise quantify $R1s$ (Assländer, Mao, et al., 2024), which takes on significantly smaller values than $R1f$ (Helms & Hagberg, 2009; Manning et al., 2021; Samsonov & Field, 2021; van Gelderen et al., 2016). We hypothesized that eliminating this constraint on $R1s$ increases qMT’s sensitivity to changes in the semisolid pool’s biophysical properties arising from Aβ accumulation.

To test the above hypothesis, we compare the unconstrained qMT parameters with their equivalents in the constrained qMT model. Specifically, we compute the apparent macromolecular pool size $m0s,a$ as

the apparent magnetization exchange rate $Rxa$ as

and the apparent longitudinal relaxation rate $R1f,a$ as

Full details regarding these expressions for the constrained qMT model can be found in Section 2.2 of Assländer, Mao, et al. (2024).

We recruited 15 cognitively normal participants (with a Clinical Dementia Rating^® of 0) from New York University’s AD Research Center (ADRC) cohort of community-dwelling elderly adults. The assessment of being cognitively normal was based on a comprehensive set of psychometric tests across multiple cognitive domains that informed a consensus diagnosis. Hence, the yearly conversion rate to MCI is extremely low in this cohort within the NYU ADRC. All 15 participants had received amyloid PET scans within the preceding 34 months (mean $±$ standard deviation 17 $±$ 9 months). Six individuals were considered Aβ$−$ by our ADRC’s standardized uptake value ratio (SUVR) threshold (see Section 2.8), with the following demographics: three female, three white, age 72.6 $±$ 4.5 years (mean $±$ standard deviation), Montreal Cognitive Assessment (MoCA) scores 27.3 $±$ 1.2, one APOE-ε4 and two APOE-ε2 carriers. The remaining nine participants were Aβ$+$⁠, with the following demographics: five female, seven white, age 75.6 $±$ 6.5 years, MoCA 27 $±$ 2.2, and six APOE-ε4 carriers. All MoCA scores were taken from each participant’s annual ADRC visit that was closest in time to the date of the qMT scan. With the exception of one Aβ$+$ participant, all MoCA scores were obtained within a 10-month period either preceding or following the respective qMT scan. All participants provided written informed consent for the studies described below, in agreement with the requirements of the New York University School of Medicine Institutional Review Board.

All participants received 300 MBq (8.1 mCi) of [¹⁸F]florbetaben (FBB) tracer (Life Molecular Imaging, Totowa, NJ) intravenously over 15 s, followed by a 12-cc saline flush. Syringes were assayed pre- and postinjection. Participants rested in the injection room to achieve brain equilibration of the tracer before being positioned in the scanner. Amyloid PET scans were acquired on a 3 Tesla Biograph mMR PET/MRI system (Siemens Healthineers, Germany) 90–120 min postinjection. Structural MRI was also obtained for registration to images from the next session.

In a second session 17 $±$ 9 months (no more than 34 months) later, each participant underwent a 24-min MRI examination on a 3 Tesla Prisma MRI scanner (Siemens Healthineers, Germany) using a 32-channel head coil. Our experimental whole-brain qMT technique used a hybrid-state sequence (Assländer et al., 2019) optimized for quantifying MT parameters with a nominal 1 mm isotropic resolution (effectively 1.24 mm isotropic accounting for 3D radial koosh-ball sampling of only the in-sphere of a 1 mm k-space cube) across a 256 $×$ 256 $×$ 256 mm FOV in 12 min (Assländer, Mao, et al., 2024). The hybrid-state sequence is similar to an inversion-recovery balanced steady-state free precession (bSSFP) sequence (Bieri & Scheffler, 2013; Carr, 1958) in that it uses fully balanced gradient moments per repetition time, but it also incorporates a smoothly varying flip angle and RF pulse duration between repetition times (Assländer et al., 2019) that was optimized for the encoding of the qMT parameters (Assländer, Mao, et al., 2024). However, we note that the RF pattern was optimized for measuring demyelination in white matter rather than tailored specifically for the detection of Aβ in the cortex. k-Space readout was performed with a radial koosh ball trajectory, where the direction of the 3D radial spokes was distributed across the unit sphere using a 2D golden means pattern (Chan et al., 2009), reshuffled to minimize eddy current artifacts (Flassbeck et al., 2024). More details about the hybrid-state sequence—including the encoding mechanisms and its numerical optimization for quantifying the unconstrained qMT parameters—can be found in Assländer, Mao, et al. (2024).

3D Magnetization-Prepared Rapid Acquisition Gradient-Recalled Echo (MPRAGE) (Brant-Zawadzki et al., 1992; Mugler & Brookeman, 1990) and T2-weighted Fluid-Attenuated Inversion-Recovery (FLAIR) (Hajnal et al., 1992) images were also acquired with 1 mm isotropic resolution. Both sequences were GRAPPA 2x accelerated, where the MPRAGE used a TE/TR/TI (echo time, repetition time, and inversion time) of 2.98 ms/2.3 s/900 ms for 5 min 30 s of scan time and the FLAIR used a TE/TR/TI of 392 ms/5 s/1.8 s for 5 min 57 s of scan time.

PET reconstruction used the Standardized Centralized Alzheimer’s & Related Dementias Neuroimaging (SCAN) parameters (https://scan.naccdata.org), except modified to use only a single frame of the first 10 min of data to reduce motion-related artifacts (Koesters et al., 2016): OSEM-3D (Erdogan & Fessler, 1999) with 4 iterations and 21 subsets; 344 $×$ 344 $×$ 127 grid; 2.0 zoom (1.04313 $×$1.04313 $×$ 2.03125 mm voxels); all corrections on; postreconstruction smoothing with a 2 mm full width at half maximum Gaussian kernel. Attenuation correction was performed using MR-based hybrid segmentation and atlas-based algorithm that combines tissue segmentation from a Dixon µ-map with a superimposed, coregistered, skull atlas-derived bone compartment (Koesters et al., 2016).

For the qMT sequence, we performed self-navigated motion correction retrospectively with a temporal resolution of 4 s based on low-resolution reconstructions (4 mm isotropic) in a subspace optimized for contrast between the parenchyma and CSF (Flassbeck et al., 2024; Kurzawski et al., 2020). From the corrected k-space data, we reconstructed 15 coefficient images corresponding to a low-rank representation of the MRI signal’s temporal dynamics using a subspace modeling approach (Assländer et al., 2018; Liang, 2007; McGivney et al., 2014; Tamir et al., 2017). The subspace was spanned by singular vectors computed from a dictionary of signals (or fingerprints) and, additionally, their orthogonalized gradients, which maximizes the information needed to stably perform parameter estimation for complex pulse sequences (Mao, Flassbeck, Gultekin, & Assländer, 2024). The reconstruction also utilized sensitivity encoding (Pruessmann et al., 2001; Sodickson & Manning, 1997), with coil sensitivities estimated using ESPIRiT (Uecker et al., 2014), and locally low-rank regularization to minimize undersampling artifacts and noise (Lustig et al., 2007; Trzasko & Manduca, 2011; T. Zhang et al., 2015). The reconstruction was implemented using the optimal iterative soft thresholding algorithm (Jang et al., 2023) in Julia based on publicly available source code (see Section 5). More details about the reconstruction can be found in Tamir et al. (2017) and Assländer et al. (2018).

Following image reconstruction, we estimated maps of the six qMT parameters (⁠$m0s,R1f,R2f,Rx,R1s$⁠, and $T2s$⁠) by voxel-wise fitting the unconstrained qMT model to the reconstructed coefficient images. For computational efficiency, we used a neural network-based estimator. This network, tailored for our sequence, takes the 15 complex-valued coefficients, split into real and imaginary parts, as inputs. The architecture was similar to the design described in Figure 2 of X. Zhang et al. (2022): 11 fully connected layers (with layer widths $30→256→512→1024→768→$$512→384→256→128→30→32→6$⁠) with batch normalization, skip connections, and Rectified Linear Unit activation functions (see Section 5 for code; however, we note that the specific architecture employed is not essential in determining the neural network’s performance). The network incorporated a data-driven correction for $B0$ and $B1+$ inhomogeneities as described in Assländer, Gultekin, et al. (2024), and was trained explicitly to minimize the bias that is typically introduced when assuming a distribution for the simulated training data (Mao, Flassbeck, & Assländer, 2024). Example parameter maps for an Aβ$+$ participant are shown in Figure 2.

The MPRAGE, FLAIR, and FBB images were skull stripped (Hoopes et al., 2022) and rigid-body registered (Reuter et al., 2010) to the estimated $m0s$ maps (which has the most similar contrast to the MPRAGE among the qMT parameters) using the “mri_synthstrip” and “robust_register” commands, respectively, in FreeSurfer v7.4.1 (Fischl, 2012). The qMT parameters were chosen as the common reference frame to avoid interpolating the nonlinearly processed qMT maps. Cortical and subcortical segmentation, cortical parcellation, volume, and thickness values were then computed using FreeSurfer’s “recon-all” command (Fischl, 2004; Fischl et al., 2002). Volumes were normalized by the estimated total intracranial volume, and the average global cerebellar FBB value was used for normalization to compute SUVRs.

Because the cortices are only 2–3 mm thick compared with the 1.24 mm effective isotropic resolution, and since the quantification of the several qMT parameters is unstable in the CSF (where the macromolecular pool size $m0s$ approaches zero), cortical analyses of our qMT parameter maps are highly susceptible to partial volume effects. To mitigate this issue, we adopt a conservative approach of sampling the qMT and FBB SUVR values at the surface corresponding to 50% of the cortical depth using FreeSurfer’s “mri_vol2surf” function. To avoid resampling the nonlinearly processed qMT data and minimize linear interpolation error when using the FreeSurfer tools, we applied the following procedure. First, we sinc interpolated (Oppenheim & Willsky, 1996) the reconstructed coefficient images onto a 5 $×$ finer grid (i.e., 0.2 $×$ 0.2 $×$0.2 mm voxels). Then, we used “mri_vol2surf” with trilinear interpolation to sample the coefficient images at 50% of the cortical depth before applying the neural network to estimate the qMT parameter maps on the interpolated surface. We note that this approach still depends on the accuracy of FreeSurfer’s estimated gray and white matter surfaces.

The amyloid SUVR positivity threshold of 1.08 was determined by the NYU AD Research Center through an image processing procedure on their entire cohort of participants who have received florbetaben PET scans since 2021, independent of the analyses performed in this manuscript. According to the recommended ADNI methods (Landau et al., 2021), FreeSurfer-derived Desikan–Killiany atlas regions were used to construct a composite neocortical ROI comprising the frontal, lateral parietal, and lateral temporal cortex as the “target” region, where the whole cerebellum served as the “reference” region. The neocortical SUVR was computed based on the mean activity within.

Due to small sample sizes, we perform a lobar-level analysis by grouping the Desikan–Killiany cortical ROIs (Desikan et al., 2006) into the four primary cortical lobes (frontal, parietal, temporal, and occipital) using the schema suggested in the Appendix of Klein and Tourville (2012). Each participant’s average measurement (qMT parameter, FBB SUVR, or thickness) per cortical lobe and hemisphere is estimated from the constituent ROIs using an inverse-variance weighting (Kay, 1993) based on each ROI’s sample variance as computed by FreeSurfer’s “mri_segstats” function (i.e., the maximum likelihood estimate assuming the “mri_vol2surf” samples from each cortical ROI are normally distributed). Prior to all statistical analyses, we manually excluded four ROIs (for all measurements) where we identified substantial artifacts in $Rx$ across multiple participants, likely caused by subcutaneous fat (see Section 4.2): the rostral anterior cingulate, precentral, postcentral, and superior parietal gyri. We also excluded three ROIs (for all measurements) superior to the frontal sinus that exhibited bSSFP-like banding artifacts (Bieri & Scheffler, 2013) in the qMT parameters (Fig. 2L): the medial and lateral orbitofrontal cortices and temporal pole.

We also consider a synthetic lobe based on the “signature of AD-related cortical thinning” described in Dickerson et al. (2011), which defines a group of cortical regions identified as being the most vulnerable to thinning in a large cohort of cognitively normal individuals who later developed AD dementia. From the Desikan–Killiany atlas, we chose the following ROIs that most closely matched those described in Dickerson et al. (2011) to compute the AD-signature summary measure: the entorhinal cortex, temporal pole, inferior temporal gyrus, angular gyrus, supramarginal gyrus, superior parietal cortex, precuneus, middle frontal gyrus, and superior frontal gyrus.

As a secondary outcome of our study, we also computed average measurements in subcortical structures based on the FreeSurfer segmentation: the hippocampus, amygdala, thalamus, caudate, putamen, and pallidum.

We used the nonparametric Mann–Whitney $U$ test (Fay & Proschan, 2010) to compare measurements between the Aβ$−$ and Aβ$+$ groups. We considered the $p<0.05$⁠, Bonferroni-corrected $p<0.01$ (accounting for the five cortical “lobes” considered), and $p<0.007$ (accounting for six subcortical and global WM ROIs) significance levels, where the latter two were used to account for the family-wise error rate (Dunn, 1961) in the group comparisons shown in Figures 3–5. Effect sizes were quantified using Hedge’s $g$ (Hedges, 1981), where we considered $0≤|g|<0.5$ “small,” $0.5≤|g|<0.8$ “medium,” $0.8≤|g|<1.2$ “large,” and $|g|≥1.2$ “very large” (Sawilowsky, 2009). A positive effect size (⁠$g>0$⁠) was defined as measurements that were larger in the Aβ$+$ group. We used Pearson’s correlation coefficient to evaluate the association of our measurements with amyloid burden, again using the $p<0.05$ significance level. All statistical analyses were performed using the HypothesisTests.jl Julia package.

Figure 3 compares the qMT parameters, FBB SUVR, and cortical thickness values between the Aβ$−$ and Aβ$+$ groups. Uniformly across the neocortex, SUVR values were significantly increased in Aβ$+$ participants with very large effect sizes, which was expected from our definition of amyloid positivity. Consistently with previous literature reports, the magnetization exchange rate $Rx$ was significantly decreased in the temporal lobe with a large effect size (⁠$g=−0.86$⁠). The macromolecular pool’s longitudinal relaxation rate $R1s$ was also significantly decreased in the frontal, parietal, and temporal lobes with medium (⁠$g=−0.78$⁠), large (⁠$g=−0.81$⁠), and large (⁠$g=−0.94$⁠) effect sizes, respectively. By contrast, no significant group differences were observed in any lobe for the macromolecular pool size $m0s$⁠, cortical thickness (including for the “AD cortical signature measure”), or the remaining qMT parameters (shown in Supporting Table S2).

Figure 4 shows the group comparison using the constrained qMT model described in Section 2.2. Here, the apparent magnetization exchange rate $Rxa$ is still significantly decreased in the temporal lobe, which is consistent with previous literature reports that also assume constraints on the value of $R1s$ (Giulietti et al., 2012; Makovac et al., 2018). However, $R1f,a$ does not exhibit significant group differences in any cortical lobe, unlike the unconstrained $R1s$⁠. This analysis suggests increased information coupling between the qMT parameters in the constrained model that reduces the overall sensitivity to amyloid burden.

We repeated the analysis using the unconstrained model for the FreeSurfer-defined global white matter and subcortical structures, shown in Figure 5 and Supporting Tables S3–S4. There were no significant differences in the volumes of any subcortical structures (including the hippocampus (Sabuncu et al., 2011)) or total CSF (not shown). However, we observed significant decreases in both free pool relaxation times (⁠$R1f$ and $R2f$⁠) in the global white matter, hippocampus, and thalamus. $m0s$ and $R1f$ were also significantly decreased in the pallidum.

Figure 6 analyzes the correlation between the unconstrained qMT parameters or cortical thickness and amyloid concentration (as measured by FBB PET SUVR). Significant Pearson’s correlations occur in locations similar to Figure 3, though with a few differences: there was no significant negative correlation with $R1s$ in the parietal lobe, but a significant positive correlation with $m0s$ in the temporal lobe. The joint consistency between Figures 3 and 6 improves the overall confidence in the results. For example, while the significant decrease in $Rx$ within the temporal lobe in Figure 3 appears to be driven primarily by a couple of participants, Figure 6 demonstrates that there is also a significant trend toward decreased values as a function of amyloid concentration. Similarly to Figure 3, no significant correlations were observed between cortical thickness and amyloid concentration in any lobe (neither for the “AD cortical signature measure”). Overall, the direction of correlation (positive or negative) for each measure is the same across all lobes except in the occipital lobe for $Rx$⁠. We note, however, that this outlier is not statistically significant.

To understand the spatial correspondence between the unconstrained qMT measurements and FBB SUVR, we visualize the corresponding effect sizes for each cortical ROI overlaid on the Desikan–Killiany atlas in Figure 7. Note that because increases in FBB SUVR as opposed to decreases in $Rx$/$R1s$ were observed in the Aβ$+$ group in Figures 3 and 6, a reversed color bar is used for FBB SUVR for ease of comparison across the different measurements. As expected, very large positive effect sizes are uniformly observed across the cortex for FBB SUVR. While we do not observe uniformly large negative effect sizes for $Rx$ and $R1s$⁠, the ROIs exhibiting (likely spurious) positive effects somewhat overlap with the ROIs excluded from the lobar analyses for exhibiting imaging artifacts (e.g., the postcentral and superior parietal gyri in $Rx$⁠). Importantly, though we do observe small (⁠$g=−0.42$⁠) to medium (⁠$g=−0.69$⁠) effects suggesting subtle thinning of the right/left entorhinal cortices (Dickerson et al., 2011; Sabuncu et al., 2011), the qMT parameters exhibit higher overall sensitivity to amyloid burden across the entire neocortex.

Our study revealed widespread group differences between Aβ$+$ and Aβ$−$ individuals across the various unconstrained qMT parameters. The largest effect was observed in the semisolid pool’s longitudinal relaxation rate $R1s$ which, notably, is not detectable in any of the constrained qMT parameters (cf. Fig. 4). $R1s$ was previously considered inaccessible and typically constrained. However, using the hybrid-state’s enhanced signal encoding capabilities (Assländer, Gultekin, et al., 2024; Assländer, Mao, et al., 2024; Assländer et al., 2019), we are able to quantify $R1s$ on a voxel-by-voxel basis. Our data suggest that $R1s$ may be a potential biomarker for amyloid pathology preceding morphometric measures of atrophy (Dickerson et al., 2011; Sabuncu et al., 2011). A potential biophysical explanation for this finding could be the slower longitudinal relaxation of spins bound specifically in Aβ plaques relative to other constituents of the macromolecular pool, such as myelin.

We also observed a reduction in $Rx$ in Aβ$+$ participants, which has previously been attributed to the hydrophobicity of Aβ plaques (Chen et al., 2017; Giulietti et al., 2012). This finding aligns with previous reports that the forward exchange rate (i.e., $m0s⋅Rx$⁠) is reduced in clinical AD and is predictive of amnestic MCI to AD conversion (Duan et al., 2022; Giulietti et al., 2012; Makovac et al., 2018). Our results using an unconstrained qMT model, however, suggest that $Rx$ is more sensitive to amyloid pathology than $m0s$ in preclinical AD. This finding raises an interesting possibility: if neurodegeneration causes reduced $m0s$ in advanced disease (discussed further in Section 4.2), $Rx$ and $m0s$ may each potentially be associated with the “A” and “N” axes of the A/T/N framework (Jack et al., 2018), respectively. This emphasizes the importance of separating these parameters in the unconstrained qMT model.

Contrary to our expectations of an increase in the macromolecular pool size $m0s$ corresponding with greater amyloid burden, we found no significant group differences in $m0s$⁠, although a positive correlation was observed in the temporal lobe. One explanation could be a competing effect causing a simultaneous decrease in $m0s$⁠, possibly due to concomitant neurodegeneration (preceding a macroscopic change in cortical thickness) or a change in the interstitial load of water (for example, due to reduced fluid clearance (Tarasoff-Conway et al., 2015) or vascular leakage from reactive astrogliosis (Nakahara et al., 1999)). However, the latter hypothesis appears less likely given the lack of significant changes we would expect in the free water pool’s relaxation rates $R1f$ and $R2f$ (Assländer, Mao, et al., 2024; Stanisz et al., 2004).

Our preliminary study was designed to assess the utility of unconstrained qMT biomarkers for detecting Aβ accumulation and has several limitations. Firstly, a larger cohort is needed to verify our findings, control for the effect of nuisance variables (e.g., age, sex, race), and increase the statistical power to perform ROI or voxel-level analyses. Secondly, the temporal delay between the acquisition of the amyloid PET and qMT scans introduces some bias which limits the power of our study to detect group differences. While Supporting Figure S1 shows that there is no significant correlation between the interscan time delay and the global cortical amyloid SUVR values, some biases may be present in the data as a result of this non-negligible temporal delay. Lastly, while our study revealed statistically significant differences in qMT parameters on a group level, the significant overlap in these measures between Aβ$−$ and Aβ$+$ groups means that qMT biomarkers are currently not as sensitive as amyloid PET measures for binary signal detection tasks on an individual level.

As the use of an unconstrained qMT model deviates significantly from the prior literature, our study was designed to identify the MT parameters that most closely reflect the spatial distribution of fibrillar amyloid in neocortical regions as compared with PET. Notably, our data suggest that $R1s$ may be a promising parameter for this purpose. Our proof-of-concept study was based on a prototype qMT sequence originally designed for the study of white matter. Future work will involve optimizing a sequence for the quantification of the Aβ burden in gray matter using the same procedure described in Assländer, Mao, et al. (2024). Supporting Figure S2 shows that a pulse sequence optimized for further improved SNR in all six core qMT parameters may be a promising avenue toward eliminating artifacts and improving qMT’s overall sensitivity to the Aβ plaque burden. Additional improvements to the sequence are also needed to further reduce the scan time for routine clinical use. Further, the qMT maps exhibited artifacts particularly affecting ROIs in the frontal and parietal lobes, which could explain their relative lack of significant effects compared with the temporal lobe. One likely source of artifacts in the frontal and parietal lobes is subcutaneous fat. Future work will involve correcting for the chemical shift-related artifacts to obviate the need for excluding ROIs in the analysis.

The qMT sequence has an effective 1.24 mm isotropic resolution. Future work will explore the potential advantages of this high resolution (as compared with PET) in the study of finer structures relevant to AD, including the hippocampus, cortex, and brainstem. For the latter, decreased neuromelanin and degeneration of the locus coeruleus in the pons have been associated with increased levels of CSF Aβ (Betts et al., 2019) and vulnerability to the occurrence of neurofibrillary tangles (Braak et al., 2011) in early stage AD. As suggested by Trujillo et al. (2019), qMT could potentially be used to detect decreased neuromelanin in the locus coeruleus.

Intriguingly, our data in the subcortical gray matter show a significant decrease in both of the free pool’s relaxation rates $R1f$ and $R2f$ for Aβ$+$ participants, specifically in the hippocampus and thalamus (Fig. 5 and Supporting Table S3). One possible explanation for these differences is inflammation (Leng & Edison, 2021), which is known to cause decreases in the relaxation rates (Stanisz et al., 2004). Iron is known to accumulate and colocalize with Aβ in the hippocampal subiculum based on postmortem AD tissue samples (Madsen et al., 2020; Smith et al., 1997; Zeineh et al., 2015), and concentrations of the ferrous form, which causes oxidative stress, have been proposed to increase with microglial-driven inflammation (Tran et al., 2022), which may link Aβ pathology with subsequent neurodegeneration (Leng & Edison, 2021). However, our previous data show that both $R1f$ and $R2f$ in the unconstrained qMT model are highly positively correlated with iron concentration (Assländer, Mao, et al., 2024). Future studies could incorporate quantitative susceptibility-weighted imaging (Ayton et al., 2017; Yedavalli et al., 2021) to verify possible hippocampal changes in iron content within the preclinical AD population.

Unconstrained qMT can also be used to study injury in white matter areas (Makovac et al., 2018), where $m0s$ was previously demonstrated to be closely associated with myelin content (Janve et al., 2013; Thiessen et al., 2013), but amyloid PET has significant off-target binding (Chapleau et al., 2022). Additionally, we hypothesize that cortical gray matter neurodegeneration in clinical AD would also be reflected by reduced $m0s$⁠. Under the model of amyloid accumulation and progressive neurodegeneration as temporally displaced processes (Jack et al., 2010), this motivates further investigation into unconstrained qMT’s sensitivity to patients with more advanced disease and specific associations with the neurodegenerative axis of the A/T/N framework. Specifically, qMT’s potential sensitivity to both the “A” and “N” axes (Jack et al., 2018) might improve the specificity for and the monitoring of progression between preclinical, MCI, and dementia stages of AD. Along similar lines, more work is also needed to clarify whether qMT parameters are sensitive to aggregates of tau proteins (like neurofibrillary tangles).

Mechanistic studies are needed to elucidate the specific pathophysiological processes underpinning the observed changes in $Rx$ and $R1s$⁠, including potential covariates such as blood–brain barrier breakdown leading to changes in perfusion (Tarasoff-Conway et al., 2015), temperature differences (Birkl et al., 2013; Klegeris et al., 2006), or pH changes due to, for example, mitochondrial dysfunction (Louie et al., 2008; Wang et al., 2020). Additionally, a postmortem study of the relationship between $Rx$/$R1s$ and neuritic plaque density would provide a stronger histopathological basis for our data. Lastly, our findings should be validated in preclinical (e.g., mouse) models of AD, as our unconstrained model differs from the existing literature primarily utilizing constrained qMT approaches (Bigot et al., 2014; Pérez-Torres et al., 2014; Praet et al., 2016).

While amyloid PET is a well-established technique with a simpler and more robust processing pipeline that offers much larger effect sizes, qMT—which combines high-resolution anatomical imaging and amyloid sensitivity in a single examination—could potentially be more amenable to screening, longitudinal and multicenter imaging studies by virtue of being implementable on existing MRI scanners. However, future work is still needed to establish the repeatability and reproducibility of the unconstrained qMT parameters, as well as streamlining the image reconstruction and model fitting steps to reduce the technical complexity. If these can be achieved, qMT imaging could potentially be useful for monitoring the longitudinal response to disease-modifying drugs in emerging clinical trials of antiamyloid immunotherapeutics, which are increasingly being studied in preclinical patient populations (Yadollahikhales & Rojas, 2023); e.g., in the AHEAD 3-45 study (NCT04468659).

Our study is the first to utilize an unconstrained qMT model to compare qMT parameters directly with an accepted measure of amyloid burden (amyloid PET) in presymptomatic individuals on the Alzheimer’s disease spectrum. We show that the magnetization exchange rate and the semisolid spin pool’s longitudinal relaxation rate are potential biomarkers of amyloid beta accumulation. While it does not achieve the sensitivity of amyloid PET, qMT is an augmentation to routinely used conventional MRI that may enable the detection of amyloid accumulation without requiring contrast agents or radiotracers. Future studies are needed to establish a potential role for qMT in monitoring disease progression and the response to therapy.

The qMT parameter maps, MPRAGE, and FBB PET SUV images for all participants are available at https://osf.io/d6r4h/ (doi: 10.17605/OSF.IO/D6R4H). The source image reconstruction code used for the qMT data is available as a Julia package on Github at https://github.com/JakobAsslaender/MRFingerprintingRecon.jl. For the presented data, we used package v0.7.0 with Julia v1.10.0. Julia code to train the neural network used for qMT model fitting is also available on Github at https://github.com/andrewwmao/BiasReducedNetworks.

A.M.: Conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, project administration, software, validation, visualization, writing—original draft, and writing—review & editing. S.F.: Methodology, software, and writing—review/editing. E.M.: Software and writing—review/editing. A.V.M.: Funding acquisition, resources, and writing—review & editing. H.R.: Methodology, resources, supervision, and writing—review & editing. J.A.: Conceptualization, funding acquisition, methodology, software, supervision, and writing—review & editing.

This work was supported by NIH grants F30 AG077794, R01 NS131948, T32 GM136573, an NIA-funded Alzheimer’s Disease Research Center (P30 AG066512), and was performed under the rubric of the Center for Advanced Imaging Innovation and Research (CAI2R), an NIBIB National Center for Biomedical Imaging and Bioengineering (P41 EB017183).

The authors declare no competing interests.

The authors are grateful to David H. Salat, Ryn Flaherty, and Yu Veronica Sui for their helpful pointers regarding the use of the FreeSurfer package, as well as Zena Rockowitz for her assistance with participant recruitment.

Supplementary material for this article is available with the online version here: https://doi.org/10.1162/imag_a_00367.