Effective Modeling of Continuous Wave Z‐Spectra for Quantitative CEST and CESL MRI Across Exchange Regimes Authors: Chris Lippe, Verena Hoerr Published: 2026 ABSTRACT AbstractPurposeChemical exchange saturation transfer (CEST) and chemical exchange‐sensitive spin‐lock (CESL) MRI enable spectroscopic imaging with high sensitivity, but conventional analysis methods often fail as exchange approaches the fast‐exchange limit. We introduce the AB‐MT+n model for robust, reproducible quantification of continuous‐wave Z‐spectra across exchange regimes.Theory and MethodsThe AB‐MT+n model describes continuous‐wave Z‐spectra by a five‐parameter basis for direct water saturation, semi‐solid magnetization transfer, and fast exchange, which can be extended by off‐resonant components for separable slow‐ and intermediate‐exchanging pools. Performance was evaluated in simulations, phantoms, and in vivo glucose‐enhanced CESL (glucoCESL) of a mouse brain.ResultsSimulations showed that AB‐MT+n recovers the maximally accessible information from Z‐spectra with minimal parameterization, with lower model orders sufficing for faster exchange. In phantoms, the slower‐exchanging glutamate pool was quantified reliably by Lorentzian analysis (R2>95%), whereas fast‐exchanging glucose was not captured accurately by Lorentzian, MTR, or the full analytical model under varying glutamate concentration (R2<45%). In contrast, the pool‐unspecific asymmetry parameter A preserved strong linearity with glucose concentration (R2>80%). Reduced model orders also sufficed in seven‐pool gray matter simulations and enabled robust quantification of in vivo glucoCESL data.ConclusionAB‐MT+n provides an accurate, acquisition‐independent, modular description of continuous‐wave Z‐spectra and improves quantification, particularly in the fast‐exchange regime.Keywords: Bloch‐McConnell equations, CEST, quantitative MRI, spin‐lock Introduction Chemical exchange saturation transfer (CEST) MRI [1] and chemical exchange‐sensitive spin‐lock (CESL) MRI [2] are powerful spectroscopic imaging techniques that leverage chemical exchange of labile protons for contrast enhancement. Studies have demonstrated a wide range of potential applications in which chemical exchange (CE) allows for detection of metabolites at millimolar concentration [3, 4]. Nevertheless, various challenges regarding the quantification of CEST/CESL MRI data persist, including long acquisition times, motion artifacts, sensitivity to B0‐, B1‐ and T1‐inhomogeneities, and low spectral resolution/selectivity [5, 6]. In general, two different quantification paradigms for CEST/CESL MRI data exist: metric‐based [7] and model‐based [8] analysis. Often employed metrics are magnetization transfer ratio asymmetry as well as magnetization transfer ratio Rex and apparent exchange‐dependent relaxation which require a labeled Z value Zlab that includes exchange effects and one reference measurement Zref without exchange effects. For MTRasym, the reference signal is approximated using the signal at the opposite frequency offset. AREX additionally requires T1 to be provided. These metrics are usually easy‐to‐calculate and robust, but have limited capabilities of correcting for field inhomogeneities and separation of individual CE effects. Model‐based approaches have potential to overcome these limitations and also allow for extracting of physically meaningful quantities like proton fraction, exchange and relaxation rates. However, for in vivo Z‐spectroscopy, models often become very complex due to the large amount of involved exchanging pools, impairing stability and reproducibility of fit results. This work aims to investigate and develop further existing model‐based analysis techniques for CEST/CESL MRI towards robust and reproducible quantification of Z‐spectra, with a particular focus on fast‐exchanging pools. All models used for Z‐spectra quantification originate from Bloch‐McConnell (BM) equations [9], which extend Bloch equations by the exchange of magnetization between multiple pools, for example through CE of labile protons. BM equations can be numerically solved for a series of irradiation offsets Δω to predict the Z‐spectrum of a multi‐pool system. Alternatively, analytical solutions exist that require approximations to different amounts. The most general model we consider here is the approximate solution for the Z‐spectrum under continuous wave (cw) irradiation proposed by Zaiss et al. [10, 11], based on the work of Trott and Palmer [12, 13]. Although, cw irradiation is not applicable in clinical settings due to specific absorption rate (SAR) restrictions, models originally derived under cw assumptions are widely used also on human CEST MRI data [14, 15]. Also, there are large efforts made to restore cw spectra properties using pulsed saturation schemes [16]. This work introduces the AB‐MT+n model, which transforms the analytical model from Zaiss et al. into a modular expression of a five parameter basis set to describe global effects of direct water saturation (DS), magnetization transfer (MT) of semi‐solid proteins and fast‐exchanging pools, extensible for sharp‐resonant effects of slow‐ to intermediate‐exchanging pools. First, we demonstrated that the AB‐MT+n model can recover the maximally accessible information from Z‐spectra in a simulated four‐pool toy‐model, and we investigated how this recoverable information depends on the simulation parameters. Second, we systematically compared the quantification performance of different models, parameters, and metrics, including AB‐MT+n, in dedicated phantom experiments. Finally, we demonstrated the applicability of the AB‐MT+n model under in vivo‐like conditions using a simulated gray matter (GM) system and measured glucose enhanced CESL (glucoCESL) data from a mouse brain. Theory Analytical Solutions to Bloch‐McConnell Equations for Quantitative Z‐Spectroscopy For long saturation times tsat>>T2, and sufficiently strong saturation ω1=γ·B1>>1/T2, BM equations are approximately solved by a one‐dimensional, mono‐exponential decay along the z‐axis [10, 17]. Specifically, the normalized z‐magnetization, that is, the Z‐spectrum, is given by where Zi=|Mi|/M0 is the initial (normalized) magnetization, Δω is the frequency offset from the water resonance, Pz,eff and Pz are projection factors (Pz,eff=Pz=1 for CESL and Pz,eff=Pz=cosθ for CEST), Zss(Δω) is the steady‐state magnetization achieved for long tsat. The steady‐state Zss is given by where the residual relaxation rate in the rotating frame R1,res is related to the intrinsic longitudinal relaxation rate of water R1w by Furthermore, the rate R1ρ(Δω) can be decomposed into: where Reff describes the intrinsic (direct) water relaxation under RF irradiation: while the exchange contribution Rex,i for each pool is given by: and further: Here, δωi are the chemical shifts, fi are the proton fractions, ki are the exchange rates, and R2i are the transversal relaxation rates of the individual pools. This solution holds under the assumption of asymmetric populations (i.e., fi<<1), and negligible longitudinal relaxation rates R1i of the pools. This model shows high accuracy compared to numerical solutions of the BM equations. However, the total number of parameters to describe a multi‐pool system is only reduced by the longitudinal relaxation rates of the exchanging pools R1i. To reduce the model complexity further, we introduce the following transformation using partial fraction decomposition, which uncouples on‐ and off‐resonant contributions that are entangled as a product of two peaks in Equation (9) (originally referred to as a‐peak and b‐peak [10]): Solutions are given by: Defining one can rewrite While A(ω1) induces an asymmetry effect around the water resonance, B(ω1) controls the width of the water peak and is directly linked to T2‐shortening effects of exchange processes. Note that B(0)=R2,obs reproduces the Swift–Connick relation [18]. The AB‐MT+n Model Zaiss et al. [11] also demonstrated that an analytical solution for a MT effect can be found analogously. It requires the following modifications to previous formulas: with and When assuming that close to the water resonance the MT contribution is dominated by the DS effects, that is, one only has to consider large saturation offsets Δω>>ω1,R1,w,R2,w, and kMT>>r1,MT, the MT exchange term can be approximated by: where and Rrf,MT(Δω) is a suitable absorption line‐shape function (Lorentzian, Gaussian or Super‐Lorentzian). Equation (19) does not fulfill two properties of the original solution proposed by Zaiss et al. [11], RexMT(Δω=0)=0 and ∂ΔωRexMT(Δω=0)=0. These properties can be restored by multiplying Equation (19) with Δω2/(Δω2+ω12), while not affecting the far off‐resonant case. A visual comparison of the described approximation with the original analytical solution for RexMT can be found in Supporting Information Figure S1. This leads to the final expression for R1ρ(Δω): with We name this model AB‐MT+n, as it includes on‐resonant asymmetry (A) and line‐width (B) effects, MT effects, as well as the off‐resonant contribution of n additional pools. Compensation of Read‐Out Imperfections The dynamic model in Equation (4) does not capture two effects previously described by Sun [19] for metric‐based analysis of non‐steady‐state Z‐spectra. They can also be considered for model‐based analysis: Longitudinal relaxation between saturation and read‐out due to delays tdelay arising from required spoilers, field‐of‐view saturation or long read‐outs, reduces the labeling efficiency by a factor C=exp(−R1,obs·tdelay). Longitudinal relaxation between saturation and read‐out due to delays tdelay arising from required spoilers, field‐of‐view saturation or long read‐outs, reduces the labeling efficiency by a factor C=exp(−R1,obs·tdelay). The initial magnetization Zi varies due to residual longitudinal magnetization Zres(Δω)=Z(Δω)·1/sin2αeff−1 after excitation in gradient‐echo snapshot sequences for an effective flip‐angle αeff. The initial magnetization Zi varies due to residual longitudinal magnetization Zres(Δω)=Z(Δω)·1/sin2αeff−1 after excitation in gradient‐echo snapshot sequences for an effective flip‐angle αeff. In Z‐spectroscopy, the signal for a finite sequence of frequency offsets Δωi is measured. After signal read‐out at one offset, the z‐magnetization will recover by until the next saturation phase. That is, the dynamic model has to be solved iteratively. However, for fitting, the model only needs to be calculated at the measured offsets. Then, one can simply use the measured Z value at offset Δωi−1 to calculate Zrec(Δωi). In total, the dynamic model for acquired Z‐spectra is given by: Distinguishability of Exchanging Pools The previously introduced transformation leading to Equation (21) did not reduce the number of parameters for the CE pools. However, it showed that on‐resonant contributions do not allow for a separation of individual pools along the spectroscopic dimension, as A(ω1),B(ω1) depend solely on ω1. For single‐B1 experiments, they can be treated as constants. If two pools should be distinguishable along the spectroscopic dimension, this has to arise from the off‐resonant contributions. Consider two fast‐exchanging pools with negligible R2i, which fulfill the following conditions: Then, they can be approximately described by one effective pool with exchange rate keff and chemical shift δωeff and their proton fractions are additive feff=f1+f2. This fusion of pools reduces the number of distinguishable pools n in the sum of off‐resonant contributions. It is often implicitly assumed to group different metabolites by their predominantly exchanging functional group (e.g., hydroxyl, amine,…). In addition, for very fast exchanging pools, the off‐resonant contribution Rex,ioff is suppressed by ki2 compared to the on‐resonant contribution Rex,ion, and might be completely omitted. The AB‐MT part therefore represents a basis set of five parameters {A,B,fMT,R2,MT,δωMT} required for all Z‐spectra, which can be extended by Rex,ioff terms of n distinct observable pools, each adding three to four more parameters {δωi,fi,ki,(R2,i)} to the total model. A detailed discussion on appropriately choosing the number of effectively distinguishable pools n will follow in subsequent paragraphs. Methods Simulation of Multi‐Pool Z‐Spectra Steady‐state Z‐spectra were simulated using Pulseq‐CEST [20] and fitted with multiple analytical models to reason appropriate model selection for quantitative Z‐spectroscopy. Rician noise was added to mimic realistic conditions and test robustness of the fitting. Given a noise‐free Z‐spectrum with signal intensities, the noisy signal can be modeled by drawing samples from the Rice distribution. The probability density function (pdf) of the Rice distribution is given by [21] where I0 is the modified Bessel function of the first kind and σ represents the noise standard deviation. We employed a gradient‐based L‐BFGS‐B solver from SciPy [22, 23] to fit Z‐spectra with the analytical steady‐state solution (Equation (5)). Different models for R1ρ were inserted in this equation, including the full analytical model proposed by Zaiss et al. [10, 11] and the here proposed AB‐MT+n model with different numbers of explicit pools considered. Fit parameters were initialized with corresponding simulation values. Bounds were set to ±50% deviation from initial value. First, a four‐pool toy model was simulated, comprising one water‐pool and one MT‐pool, as well as two CE pools which share the same variable exchange rate k=k1=k2. Further, a Lorentzian MT line‐shape was assumed, and B0=9.4T and B1=1.0μT. Exchange rates varied from 200 to 20,000 Hz. Hundred different noise patterns with σ=0.01 were generated and added to the simulated spectra. Second, a GM multi‐pool model as proposed by van Zijl et al. [24] was simulated at B0=3 T and multiple saturation amplitudes (B1=0.5,1.0,2.0,4.0μT). It consists of a water pool, one MT pool with Super‐Lorentzian line‐shape, one combined relayed nuclear overhauser effect (rNOE) pool, as well as amide, amine, guanidino, hydroxyl pools. To account for the noise‐dependence of model‐order selection, 100 noisy Z‐spectra (σ=0.005) were generated for each B1‐value and fitted with the analytical AB‐MT+n model, while varying the number of explicitly modeled CEST pools (n). Pools were added in order of decreasing exchange‐time scale (fastest pools removed first when reducing n). For every fitted spectrum, BIC was calculated, and the BIC‐optimal model order was recorded. Constant simulation parameters of both models can be found in Supporting Information Table S1. To assess how well the n fitted explicit components represented the true off‐resonant exchange contributions, we additionally evaluated the isolated ground‐truth off‐resonant Rex,ioff terms of the simulated GM pools from Equations (11) and (12), We then compared these to the effectively fitted off‐resonant contributions R^ex,joff (j=1,…,n). For each fit, normalized root‐mean‐square error (NRMSE) between fitted R^ex,joff and corresponding ground‐truth Rex,ioff was calculated to quantify component recovery: Mixed Glc/Glu Phantoms All experiments were performed on a BioSpec Maxwell 9.4 T small animal MR system (Bruker BioSpin GmbH, Ettlingen, Germany) with ParaVision 360 3.6, equipped with an actively shielded gradient system (maximum gradient amplitude of 961 mT/m, slew rate = 4600 T/m/s), and a dual setup of a 82 mm transmit‐only volume resonator and a four‐element receive‐only cryogenic probe. In total, nine different phantoms were prepared: D(+)‐Glucose (Glc, VWR International) and L‐Glutamic acid (Glu, Sigma Life Science) were dissolved at three different concentrations each (10, 20, 40 mM) in phosphate buffered saline (PBS) with 1% agarose (Carl Roth) added. Room temperature (20∘C) was maintained for all measurements. Each phantom was measured with two different saturation schemes: (a) cw‐CEST saturation with tsat=4000 ms and B1=2μT, TR=5000ms, (b) cw‐CESL saturation with tsat=500 ms and B1=5μT, TR=2000 ms. 191 frequency offsets were measured: −50 ppm, −49 ppm, …, −5 ppm, −4.9 ppm, …, 4.9 ppm, 5 ppm, …, 49 ppm, 50 ppm. Spectra were normalized with a reference scan without saturation and TR=20 s. All other imaging parameters were unaltered: Gradient‐echo snapshot sequence, 40×40 image matrix, 0.5×0.5 mm resolution, 2 mm slice thickness, TE=2.9 ms, effective TE=84.7 ms. To correct for B0‐ and B1(+)‐inhomogeneities, we performed a WASABI [25] experiment, and corrected Δω and ω1 pixel‐wise. R1,obs was provided through an additionally acquired T1‐map (inversion recovery, 20 geometrically‐spaced inversion times up to 8000 ms). The obtained spectra were fitted pixel‐wise with the full analytical model [10, 11] (Equations ((4), (5), (6), (7), (8), (9), (10))), the AB‐MT+n model (Equations (25) and (21)), as well as a Lorentzian model [24]: with the peak height ai, width Γi, center δωi, and 095%), but Glc could not be quantified accurately when the Glu concentration increased (R2<45%). Similarly, the full analytical model also failed to quantify glucose accurately (R2<25%), although it does not assume any particular exchange regime. In contrast, the AB‐MT+n model could quantify fast‐exchanging Glc (R2>80%). Although explicitly considering a Glc pool in the AB‐MT+n model was beneficial to improve quantification results, the pool‐unspecific asymmetry parameter A reflected the Glc concentration best. It showed the least interference with the competing Glu concentration, although all exchanging pools additively contribute to A, by definition. This enforces the previous observation that asymmetry effects of fast‐exchanging pools are non‐localized. For non‐constant Glu concentration, quantification quality depended on the chosen saturation scheme, indicating that different weightings can be created through the A(ω1)‐dependency. Practical Application of the AB‐MT+n Model The value of the AB‐MT+n model lies in its modular nature. It allows the model order n to be chosen based on the number of distinct observable off‐resonant effects rather than the number of theoretically contributing pools. The practical advantage of this could be shown in the GM seven‐pool simulation. Although all simulated pools contributed to the measured Z‐spectra in principle, they were not necessarily separable in practice. Due to the substantial spectral overlap of rNOE, amide, amine, guanidino, and hydroxyl contributions, together with the broad MT background, the information content of the spectra was lower than the nominal number of exchanging compartments would suggest. Accordingly, the BIC favored a reduced model order, indicating that the experimentally observable complexity is smaller than the underlying biophysical complexity. This effect became more pronounced at higher B1, where broader saturation and stronger overlap further diminish the distinguishability of individual pools, such that at B1=4.0μT the AB‐MT part alone was sufficient to accurately describe the spectra. These findings support the interpretation of AB‐MT+n as an effective model whose order should be selected according to spectral resolvability rather than assumed tissue composition. Selection of optimal model order remains a manual task yet, as no straightforward decision rule exists, but it may be automatized in future using machine learning approaches. Additional advantages of the AB‐MT+n model include the feasibility of arbitrary MT line‐shapes [11], as well as explicit corrections of B0‐, B1‐ and T1‐inhomogeneities, compared to implicit correction methods, requiring multiple‐B1 acquisitions [37]. Taken together, these benefits enabled us to perform robust quantification of glucoCESL data in a mouse brain using the lowest‐order AB‐MT variant. Despite the absence of an explicit Glc pool in the model, glucose‐induced changes could be captured reproducibly through the asymmetry parameter A. The observed post‐infusion increase in A across large parts of the brain, together with a significantly positive whole‐brain mean ΔA, indicates that the AB‐MT component already provides a sufficiently sensitive effective description of the relevant off‐resonant glucose‐related contrast in vivo. At the same time, the model yields additional quantitative tissue contrasts through the A, B, and fMT maps. Importantly, these results demonstrate that accurate and biologically meaningful quantification does not necessarily require explicit modeling of all potentially contributing exchange pools. Instead, for acquisition schemes dominated by broad and overlapping contributions, a reduced effective model may be preferable because it preserves robustness and interpretability while avoiding unnecessary parameterization. Alternative quantification approaches may utilize the line‐width of the DS, here reflected by the parameter B, in line with recent reports [38] which metrify the DS width for low saturation powers. The AB‐MT part may also be used as a baseline Z‐spectrum, analogous to existing background subtraction approaches [39, 40]. Assumptions and Limitations of the AB‐MT+n Framework The AB‐MT+n framework was theoretically derived under the assumption of cw saturation and practical approximations to derive closed‐form analytical solutions to BM equations. These include sufficiently long saturation, an approximately mono‐exponential approach to steady state, small exchanging‐pool fractions (fi≪1), and negligible longitudinal relaxation of the exchanging pools. For strongly pulsed, transient, or otherwise non‐cw‐like acquisition schemes, its applicability should therefore be validated against numerical simulations or experimental calibration data. The readout correction introduced in Equation (25) extends the practical signal model but does not remove these assumptions. It accounts for relaxation losses between saturation and readout as well as incomplete recovery between consecutively sampled offsets. However, the exchange‐dependent relaxation terms themselves remain based on the approximate cw theory. A central limitation concerns the interpretation of fitted parameters. Explicit components should only be interpreted as pool‐specific when the corresponding off‐resonant contributions are sufficiently separable under the chosen acquisition conditions. This was illustrated by the additional Rexoff‐based GM analysis: retained components at low and intermediate saturation powers corresponded to separable off‐resonant spectral contributions, most robustly rNOE and amide, whereas the guanidino‐matched component showed substantially larger deviations from the ground‐truth contribution. The guanidino pool therefore represents a borderline case, as its exchange rate is higher than for amide and rNOE and its spectral distance to nearby amine and hydroxyl resonances and on‐resonant effects is limited. Thus, a retained component should not automatically be interpreted as quantitatively reliable biochemical pool recovery. Information criteria such as the BIC can indicate the number of experimentally resolvable spectral degrees of freedom, but they do not identify the biochemical origin or quantitative reliability of retained components. The choice and interpretation of n should therefore also consider residual structure, parameter stability, expected resonance positions, acquisition settings, and the biological question of interest. Further, although different MT line‐shapes can be incorporated into the AB‐MT+n framework, the present work was restricted to comparatively narrow CEST/CESL offset regimes around the water resonance; applications outside this regime would require dedicated validation. Taken together, these limitations emphasize that AB‐MT+n should be used to extract the resolvable information contained in a given cw Z‐spectrum. Further improvements require complementary measurements that add independent information about exchange, relaxation, or more advanced analysis strategies. Strategies to Overcome Current Limitations The following strategies may help resolve information that is not identifiable from cw Z‐spectra using AB‐MT+n alone: Multi‐pH calibration of glucose exchange: Zaiss et al. [26] demonstrated that using multi‐pH measurements of natural D‐glucose allows for calibration of exchange rates under physiological conditions, compared to multi‐B1 measurements only. Multi‐pH calibration of glucose exchange: Zaiss et al. [26] demonstrated that using multi‐pH measurements of natural D‐glucose allows for calibration of exchange rates under physiological conditions, compared to multi‐B1 measurements only. Prior information and multi‐contrast experiments: Incorporating independently measured relaxation and exchange rates, chemical shifts as priors/constraints stabilizes fits and reduces crosstalk [41, 42, 43]. Prior information and multi‐contrast experiments: Incorporating independently measured relaxation and exchange rates, chemical shifts as priors/constraints stabilizes fits and reduces crosstalk [41, 42, 43]. AI, fingerprinting: Dictionary‐based fingerprinting (MRF) and machine learning algorithms can map complex acquisition schedules to parameters with improved precision [44, 45, 46]. AI, fingerprinting: Dictionary‐based fingerprinting (MRF) and machine learning algorithms can map complex acquisition schedules to parameters with improved precision [44, 45, 46]. Generalized Bloch modeling: AB‐MT+n is limited to quantification of cw Z‐spectra. The generalized Bloch framework by Assländer et al. [47, 48] may offer additional insights when approaching the hard‐pulse limit. Generalized Bloch modeling: AB‐MT+n is limited to quantification of cw Z‐spectra. The generalized Bloch framework by Assländer et al. [47, 48] may offer additional insights when approaching the hard‐pulse limit. Conclusion The AB‐MT+n model provides an accurate, acquisition‐independent description of cw Z‐spectra and enables extraction of the maximally recoverable information content from Z‐spectra. The AB‐MT basis consists of five parameters that capture global contributions from DS, semi‐solid MT, and fast‐exchanging pools. This basis can be extended by the number of separably measurable pools, not leading to effective overfitting. Due to its modular structure and robustness, the AB‐MT+n framework is particularly suited for dynamic CEST/CESL acquisitions and for quantitative MRI approaches that integrate complementary readouts. Funding This research was supported by the Deutsche Forschungsgemeinschaft (German Research Foundation) (Grant Nos. CRC1450‐431460824 and 493484898). Supporting information Figure S1: Comparison of the effective water relaxation rate Reff, the original MT relaxation rate contribution RexMT, and the approximate MT expression used in the AB‐MT+n framework. Table S1: Constant simulation parameters used for the four‐pool toy model and realistic gray matter model. Table S2: Parameter bounds used for model fitting.