Phase‐Sensitive Modeling Improves Fat DESPOT Multiparametric Relaxation Mapping in Fat‐Water Mixtures Authors: Renée‐Claude Bider, Cristian Ciobanu, Jorge Campos Pazmiño, Véronique Fortier, Evan McNabb, Ives R. Levesque Published: 2026 ABSTRACT ABSTRACTPurposeTo improve on a multiparametric technique—known as Fat DESPOT—to map the fat‐ and water‐specific R1 (R1f, R1w), R2*, and proton density fat fraction (PDFF) by upgrading the parameter initialization and introducing explicit model sensitivity to the complex phase of the water and fat signals.MethodsWe compared three‐point Dixon and Graph Cut (GC) approaches to initial guesses for Fat DESPOT in experiments at 3 T in fat emulsion phantoms. Also in the phantom, we compared magnitude‐based Fat DESPOT (Fat DESPOT m ) to a magnitude approach modeling the phases of fat and water separately (Fat DESPOT mϕ ), and two models for complex data (Fat DESPOT c , Fat DESPOT cϕ ). The complex approaches, which performed best, were used in the lower leg and abdomen of healthy human participants.ResultsIn phantoms, Fat DESPOT using three‐point Dixon and GC returned similar parameter estimates and precision, though the former deviated from the overall trend at 50% nominal fat fraction. The complex models showed the best agreement with reference PDFF (average error 1.2% ± 1.2%) and the lowest combined standard deviation across ROIs, for PDFF, R1f, and R1w (maximum σ = 0.13%, 0.19 s−1, and 0.009 s−1, respectively).ConclusionWith more precise R1f and R1w, more accurate PDFF, and more versatile echo time selection, this work demonstrates the advantages of Fat DESPOT multiparametric mapping with complex fitting and GC‐based parameter initialization. Introduction Quantitative MRI mapping of proton density fat fraction (PDFF) and relaxation parameters R1 and R2* offers promising insights into disease. Notably, mapping R2* and PDFF in the pancreas is sensitive to iron content in patients with a variety of diseases [1, 2]. Meanwhile, mapping PDFF, R1, and R2*, could differentiate between types of liver disease [3, 4] and correlate with treatment outcomes [4, 5, 6]. Finally, R2* and R1 mapping is valuable in tumor hypoxia [7]. However, when separate acquisition protocols are required for each measured parameter, long acquisition times are taxing on patients, increasing the risk of motion artifacts and limiting dynamic imaging. Multiparametric mapping, where a single acquisition protocol obtains maps for several parameters, can significantly reduce acquisition times [6, 8]. Fat‐water relaxation mapping [8, 9], also referred to as fat‐water separated driven equilibrium single point observation of T1 or “Fat DESPOT” [10], is a multiparametric technique that models the signal magnitude in a variable flip angle (VFA) multi echo gradient echo (mGRE) experiment to obtain maps for PDFF, R2*, R1 of water (R1w), and R1 of fat (R1f), simultaneously. Fat DESPOT is versatile in its potential applications. The isolated R1w could be used in MRI‐based assessments of liver disease [11, 12]. R1f mapping has been proposed as a sensitive method for R1‐based MR oximetry [13] due to the increased solubility of oxygen in fat relative to water. Prior work has demonstrated the oxygenation sensitivity of Fat DESPOT in phantoms [10]. In addition, the 3D mGRE sequence required for Fat DESPOT is broadly available on clinical scanners, enhancing translatability. Four signal models were considered in this work. The magnitude model [10], referred to here as Fat DESPOT m , models the signal magnitude, assumes that the water and fat have identical initial phase, and is insensitive to variations in B0. Building on this, we tested three phase‐sensitive Fat DESPOT models. The first models the signal magnitude but incorporates a term for the initial phase difference between the water and fat signals. The other two consider the complex MR signal and explicitly model variations in B0 and in the initial phase of the water and fat signals. One complex model imposes that the initial phases are distinct but independent of the excitation pulse flip angle (FA), and other imposes no such constraint. Prior work suggests that Graph Cut (GC) fat‐water separation outperforms three‐point Dixon with region‐growing B0 estimation [14]. GC also provides R2* and the initial phases of fat and water magnetization to initialize the Fat DESPOT fit. Finally, GC offers greater flexibility of echo train design (number and spacing of echoes). We began by comparing the use of two fat‐water separation techniques, three‐point Dixon and GC, to provide initial guesses as input to the Fat DESPOT m fit. Should GC reveal itself to be equivalent (or superior) to three‐point Dixon as an initialization approach, then the added experimental flexibility with GC would be beneficial to the overall experimental design. Then, we conducted a systematic comparison of the four signal models in phantom, to assess performance across a wide range of fat fractions. Finally, we present pilot measurements in the lower leg of a human volunteer and in the abdomen of another, using the complex approaches. Methods All calculations were completed in MATLAB (Mathworks, USA, R2023a and R2024b). Signal Model This paper compares four models for Fat DESPOT mapping of PDFF, R2*, R1f, and R1w. The most general approach, Fat DESPOT cϕ (Equation 1), models the complex mGRE signal, accounting for the B0 field inhomogeneity (ΔB0,θ; fixed parameter for each FA), and for the initial phases of fat and water magnetization, ϕ0f,θ and ϕ0w,θ, free parameters that have been shown to vary experimentally and to depend on the excitation FA and on relaxation times [15]. The parameters TR, TE, and θ are the mGRE sequence repetition and echo times, and the excitation pulse FA, respectively. Fat DESPOT c (Equation 2) also models the complex mGRE signal but imposes that the initial phases of fat and water magnetization, ϕ0f and ϕ0w do not depend on FA, thereby reducing the number of free parameters. Taking the magnitude of the complex Fat DESPOT c model yields a third model, Fat DESPOT mϕ (Equation 3), where the initial phase difference between fat and water, Δϕ0=ϕ0f−ϕ0w, remains as a free parameter in the fit, independent of FA, and the ΔB0 heterogeneity drops out. Finally, Fat DESPOT m models the magnitude signal and assumes that the initial phases of fat and water magnetization are identical (ϕ0f=ϕ0w) such that Δϕ0=0 drops out of the equation, as in previous work [10]. For all models, f is the PDFF, W and F are factors that describe steady‐state relaxation and radio‐frequency pulse effects on the fat and water signal components for TE = 0 and normalized by the equilibrium magnetization (Equations 5 and 6, respectively). The fat spectrum with N resonances is represented by amplitudes An and frequency shifts Δωn. W and F are sometimes defined as including factors representing the PDFF and equilibrium magnetization (e.g., S0(1−f)W and S0fF); furthermore, S0(1−f)W and S0fF have also been denoted as ρW and ρF [14], SW and SF [15], and MW and MF [16]. To allow for B0 field map variations between acquisitions with different FA, which we observed experimentally, the parameter ΔB0,θ was introduced into the complex models as a flip‐angle specific fixed parameter. (1) Scϕ(TE,TR,θ)=S0e−R2*TEe−iγΔB0,θTE(1−f)We−iϕ0w,θ+fF∑n=1NAne−iΔωnTEe−iϕ0f,θ (2) Sc(TE,TR,θ)=S0e−R2*TEe−iγΔB0,θTE(1−f)We−iϕ0w+fF∑n=1NAne−iΔωnTEe−iϕ0f (3) Smϕ(TE,TR,θ)=S0e−R2*TE(1−f)W+fF∑n=1NAncosΔωnTE+Δϕ02+fF∑n=1NAnsinΔωnTE+Δϕ02 (4) Sm(TE,TR,θ)=S0e−R2*TE(1−f)W+fF∑n=1NAncosΔωnTE2+fF∑n=1NAnsinΔωnTE2 (5) W(TR,θ)=1−e−R1wTR1−e−R1wTRcosθsinθ (6) F(TR,θ)=1−e−R1fTR1−e−R1fTRcosθsinθ Phantom Construction Phantom experiments were carried out to assess the impact of the method used to provide initial parameter guesses, and to compare the four signal Fat DESPOT models, across a range of fat fractions. A phantom of fat‐water emulsions was constructed following a published protocol [17] with slight adjustments. To facilitate the emulsification, the surfactant polyethylene glycol sorbitan monolaurate (Tween 20, MilliporeSigma Canada Ltd.) was added to the peanut oil. Likewise, the surfactant sorbitan monooleate (span 80, MilliporeSigma Canada Ltd), and the preservative sodium benzoate (MilliporeSigma Canada Ltd.), were added to the agar solution. Additionally, gadobutrol (Gadovist, Bayer Healthcare) was added as a relaxation agent ([Gd+] = 0.2 mM) in the agar gel preparation. Each emulsion was placed in a 50 mL conical polypropylene tube (Corning 50 mL centrifuge tubes). Two additional tubes were prepared containing pure peanut oil and the agar gel preparation (with gadobutrol and surfactants), respectively, for a total of seven nominal fat volume fractions. The tubes were then suspended on a plastic and polystyrene rig placed in a cylindrical acrylic phantom container (Magphan SMR170, The Phantom Laboratories, Salem, USA), which was then filled with a solution of distilled deionized water with gadobutrol ([Gd+] = 0.3 mM) and sodium chloride (Windsor Salt Ltd) ([NaCl] = 24 mM) to approach the conductivity of human tissue [18]. Phantom Data Acquisition All measurements of the phantom were performed at room temperature in a 3 T MRI scanner (Ingenia, Philips Healthcare) using a vendor‐provided 15‐channel receive‐only head coil. The phantom was left to rest in the centre of the bore for at least 30 min before measurements to reduce flow artifacts. For Fat DESPOT measurements, a 3D mGRE sequence with monopolar readout and default gradient and RF spoiling was employed. Measurements were collected with four excitation FAs. Eight signal averages were acquired for each measurement, and parallel imaging was not used to maximize SNR. Sequence parameters used in these experiments are summarized in Table S1. A first data set was acquired to compare the impact of initial guesses provided by two fat‐water separation approaches, three‐point Dixon and GC, when used with Fat DESPOT m . Two 6‐echo (2 × 6) series were acquired at each FA using the a 3D mGRE with ΔTE = 2.4 ms and TR = 18 ms. For the first acquisition, the initial echo time (TE1) = 1.5 ms, and for the second acquisition, TE1 = 2.7 ms. These TE1s were selected such that the two acquisitions could be combined in postprocessing to create a (2 × 6) 12‐echo train with shorter apparent ΔTE (=1.2 ms), required for robust fat‐water separation using three‐point Dixon [10]. To ensure that the scanner calibrations (including receiver gains) were held constant across multiple acquisitions, the “Preparation Phases” parameter was configured such that the first acquisition in the protocol was set to “Full”, while all subsequent acquisitions were set to “Auto.” This configuration forces a complete system calibration for the first acquisition and automatically reuses the resulting calibration parameters for the following acquisitions. Additionally, to maintain uniform signal scaling across acquisitions, the “Sequence Control” scaling parameter was set to “Density Transverse Ordered Spins (DOTS)”, which standardizes the output scaling of reconstructed images across datasets. FAs were optimized to minimize R1 estimate variance considering the TR and the expected range of R1, for a set of four FAs [19, 20]. In phantoms, a range of 0.54–2.9 s−1 was used based on other fat‐water phantoms [10, 21]. Selected FAs were θ = [3°, 6°, 15°, 34°]. In the second experiment, the models were compared using data collected with an 8‐echo 3D mGRE sequence previously proposed by our group [22]. This acquisition scheme had four FAs, eight echoes, ΔTE = 1.8 ms, TE1 = 1.9 ms, and used the minimum TR = 24 ms. FAs were reoptimized for TR = 24 ms, resulting in angles θ = [3°, 7°, 17°, 39°]. Experimental VFA R1 measurements are known to be affected by B1‐induced FA variations, which were corrected throughout with a relative B1 map from the dual‐angle method [23]. B1 map data was acquired using a multi‐slice multi‐shot turbo spin‐echo (MS TSE) acquisition at two excitation FAs (60°, 120°). A separate series was acquired with a unipolar 3D mGRE sequence (TE1 = 1 ms, # of echoes = 6, ΔTE = 1.7 ms, FA = 3°) to provide a reference measurement for PDFF. All sequence parameters are noted in Table S1. In Vivo Data Acquisition Pilot measurements using Fat DESPOT c were conducted in the lower leg and abdomen of two healthy volunteers (males, Age 24 and 30, respectively) using the same 3 T MRI system as the phantom measurements. The study was approved by the Research Ethics Board of the McGill University Health Centre, and the volunteers gave informed consent. The lower leg dataset was acquired using a vendor‐supplied 8‐channel receive‐only extremity coil. The abdominal dataset was acquired using the vendor‐supplied body array receive coils (anterior and posterior, total 32 channels). Image homogeneity correction was applied with a vendor‐supplied technique that uses receive coil sensitivity maps acquired in a reference scan (“constant level appearance” or CLEAR, Phillips Healthcare), in complement to the mapping‐based FA corrections. For the lower leg dataset, the same 8‐echo acquisition protocol was used as for the second phantom experiment, excluding the PDFF reference measurement. FAs were reoptimized to the expected R1 range of human tissue, 0.56–3.33 s−1 [24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34] with TR = 24 ms. This resulted in FAs of θ = [4°, 10°, 22°, 51°] for all in vivo measurements. A smaller FOV (192.5 × 160.4 × 100 mm3) was selected, reducing acquisition time. All other parameters were preserved from the acquisition protocol for phantom measurements, including the number of averages [8] and the absence of parallel imaging. The total scan time was 22.90 min, including B1 mapping. For the abdominal dataset, sequence parameters were modified to enable data collection under breath hold and full liver coverage. For the mGRE datasets, we reduced the number of echoes, TR, and NSA, while increasing the voxel size, compared to experiments in phantom and lower leg. Parallel imaging was not used. The FAs were unchanged (4°, 10°, 22°, 51°). For B1 mapping in abdomen, we used a single‐shot (SS) TSE with parallel imaging (SENSE factor R = 1.5 anterior–posterior direction). Each mGRE and SS‐TSE series (for FA variations) was acquired in a separate breath hold. All sequence parameters for in vivo experiments are presented in Table S2. Data Processing The mGRE data underwent preprocessing steps before Fat DESPOT fitting. For the first phantom experiment, the two acquisitions taken at each FA were combined into a single data set for each FA by alternating echoes in increasing echo order. For all approaches, the relative B1 map was constructed from the dual angle MS‐ or SS‐TSE acquisition and used to scale the nominal FA for each voxel [35]. For the leg and abdomen datasets, all images were registered to the 3° mGRE acquisitions using rigid and affine registration, respectively (function imregtform). Next, initial fat water separation was performed on the lowest FA data (=3°) using either the three‐point Dixon [36] or GC algorithm [37]. Maps of PDFF were calculated from the fat‐water separated complex signal maps using a magnitude discrimination method [38], to be used as an initial guess for Fat DESPOT. In all instances, maps of the apparent R1 of the mixture, R1global, were calculated using a VFA approach with data from the first echo [10, 39]. R1f and R1w for each voxel were initialized based on the estimate of PDFF. For PDFF > 50%, R1global was used to initialize R1f and a value of 1 s−1 for R1w. For PDFF < 50%, R1global was used to initialize R1w and a value of 4 s−1 for R1f. The initial guess for R2* depended on PDFF initialization. To complement the three‐point Dixon approach, a monoexponential fit to data with TE1 = 1.5 ms, FA = 3° was used. When GC was used to initialize PDFF, the R2* initial guess was obtained directly from GC fat‐water separation, along with ϕ0f and ϕ0w. The 3°‐FA data were used for Fat DESPOT c phase initialization and the initial Δϕ0 for Fat DESPOT mϕ , and all FAs for Fat DESPOT cϕ . Finally, for Fat DESPOT c and Fat DESPOT cϕ , ΔB0,θ maps were obtained from the GC output at each FA. Following this preprocessing, the models were fit to their respective mGRE data using a nonlinear least‐squares algorithm (function lsqnonlin using the trust‐region‐reflective algorithm). For phantom experiments, a six‐resonance model of the chemical shift spectrum of peanut oil, with chemical shifts (δn=Δωn/ω0) = [0.80, 1.20, 2.00, 2.66, 4.21, 5.20] ppm and amplitudes (An) = [0.087, 0.694, 0.128, 0.004, 0.039, 0.048], was borrowed from previous experimental measurement [40]. For the in vivo measurement in leg, an eight‐resonance fat spectrum from skeletal muscle, with δn = [0.90, 1.30, 1.57, 2.03, 2.24, 2.78, 4.13, 5.30] ppm and An = [0.089, 0.598, 0.047, 0.077, 0.052, 0.011, 0.035, 0.066], was used [41]. For the measurement in abdomen, a six‐peak fat spectrum from liver with δn = [0.90, 1.30, 2.1, 2.76, 4.31, 5.3] ppm and An = [0.087, 0.693, 0.128, 0.004, 0.039, 0.048] was used [42]. Setting upper and lower bounds for PDFF to be within ±5% of the PDFF initial guess was found to improve the accuracy of PDFF output compared to the reference measurement. All other parameter bounds are displayed in Table S2. In these experiments, the B0 field map (obtained from fat‐water separation) was observed to vary between FA acquisitions, and so the FA‐specific ΔB0,θ maps were included in the fit as fixed parameters. A flowchart of the data analysis steps is presented in Figure S1. The normalized root mean squared error (nRMSE) was used as a measure of fit quality. Normalization, relative to the magnitude of the signal from the first echo with the smallest FA, was used for ease of comparison between experiments. In the phantom, a reference PDFF measurement was obtained from the 6‐echo unipolar mGRE data described above, using GC fat‐water separation. Resulting multiparametric maps have been displayed using perceptually uniform color maps recommended for quantitative maps [43]. Statistical Analysis For quantitative measurements and statistical analysis of PDFF, R2*, R1f, and R1w, regions of interest (ROIs) were selected (Figure 1). In the phantom, seven circular ROIs were manually selected to fit within the cross‐sectional area of each tube. Every ROI had the same volume (number voxels = 243). In vivo, circular ROIs were selected for the bone marrow (in the tibia), the soleus muscle, and the liver. Rectangular ROIs were selected for the subcutaneous fat. Due to the size and shape of the bone marrow and subcutaneous tissue, a geometrical ROI could not be used alone without including voxels from neighboring tissues. Hence, the ROIs for bone marrow (ROI 1) in leg and subcutaneous fat (ROI 3 in leg and ROI 2 in abdomen) were refined by including only the voxels within the geometrical shapes with PDFF > 60% and PDFF > 70%, respectively, and with R 2 ≥ 0 (from the fit) to exclude low signal voxels such as air or cortical bone. These thresholds for PDFF were selected to include the full range of PDFFs expected based on the subcutaneous fat (74% ± 13% [44]) and bone marrow (between 82% and 94% [45, 46]) reported in the literature. The thresholds were also verified versus the distribution of PDFF within the starting geometric ROI. Thresholding ensured that the voxels included in the analysis were representative of the tissue of interest. All ROIs were measured over three slices selected centrally to the imaging volume. Regions of interest of the variable fat fraction phantom (left), the lower leg of human volunteer 1 (middle), and the abdomen of human volunteer 2 (right). In the phantom, ROIs 1–7 correspond to nominal fat fractions of 0%, 5%, 25%, 50%, 60%, 75%, and 100%, respectively. In the lower leg, ROIs 1–3 correspond to tubular bone marrow, calf skeletal muscle, and subcutaneous fat. Bone marrow and subcutaneous fat voxels of interest within the ROI were selected based on a PDFF estimate > 70% and > 60%, respectively. In the abdomen, ROIs 1–2 correspond to liver parenchyma and subcutaneous fat. All ROIs were measured over 3 slices of the acquired image. The mean and standard deviation of each parameter, and the nRMSE were calculated for each ROI in phantoms and in vivo. The combined nRMSE within each ROI was obtained using Equation (7) where K is the number of voxels being combined. Statistical analysis was carried out for experiments in phantoms. Comparison of means was conducted for each parameter in the model using a two‐way ANOVA (function anova2) to assess the effects of Fat DESPOT approach, ROI (nominal fat fraction), and interactions. Parameter means were then compared using multiple pair‐wise testing with Bonferroni correction (function multcompare with CriticalValueType bonferroni, and anova1 input) and a p value of 0.05 was deemed significant. To compare standard deviations between approaches, a two‐sample F‐test for equal variance (function vartest2) was used with a significance threshold of 0.05. To compare variability across models, the combined standard deviation across ROIs was calculated for each approach following Equation (8), where N is the number of ROIs being combined, w is the sample size of the ROI, and σ is the standard deviation. (7) nRMSEcombined=∑k=1KnRMSEk2K (8) mcombined=∑n=1Nwnσn2∑n=1Nwn Results Comparing Initial Estimates From Three‐Point Dixon and GC Analysis of the 2 × 6‐echo data analyzed with the Fat DESPOT m model returned plausible parameters, regardless of the method used to initialize PDFF (three‐point Dixon or GC). This is reflected in the fairly uniform multiparametric maps (Figure 2). Fit quality was similar across all ROIs, with a combined nRMSE of 0.20 using either approach. The highest nRMSE was 0.38 using the Dixon approach and 0.37 using the GC approach, in the ROI with a nominal fat fraction of 50%. Multiparametric maps for PDFF, R2*, R1f, R1w, and nRMSE using the three‐point Dixon and GC as PDFF initial guess inputs for Fat DESPOT m on a 2 × 6‐echo dataset. To reduce noise in the R1f images, voxels with PDFF < 3% and in the R1w images, voxels PDFF > 97% were masked. In the R2* maps, variations appeared in the large water compartment of the phantom surrounding the emulsion tubes, likely due to intra‐voxel B0 field inhomogeneity from the polystyrene support in the phantom inherently impacting the R2* regardless of the choice of Dixon or GC. Parameter values from the voxel‐wise Fat DESPOT m fit are displayed in Figure 3, extracted from ROIs shown in Figure 1. The choice of Dixon or GC fat‐water separation to initialize PDFF for Fat DESPOT m returned relatively stable and similar estimates of PDFF, R2*, R1f, and R1w across fat fractions. Average percent differences between Fat DESPOT model parameters from fits initialized with Dixon and GC were 11.6% ± 1.3% for R1f and 11.9% ± 2.3% for R1w, respectively. The most obvious impact was in the ROI with nominal fat fraction = 50%, where values calculated using the Dixon approach deviated considerably from the overall trend. Box plot of voxel‐wise estimates of PDFF, R2*, R1f, and R1w, using the three‐point Dixon and GC as PDFF initial guess inputs for Fat DESPOT m on a 2 × 6‐echo dataset. Box = interquartile range, horizontal line = median, whiskers = data range (=1.5× interquartile range), dots = outliers. Two‐way ANOVA conducted for each parameter with effects “ROI”, “Approach” (Dixon or GC for initial guesses), and interactions, revealed a statistically significant effect of fitting approach for every parameter (p<0.001). Post hoc testing confirmed the statistically significant effect of fitting approach on R1f and R1w across all but one ROI each. Conversely, PDFF and R2* estimates were only significantly different in the ROI with 50% nominal fat fraction (p<0.001). When compared to the reference measurement, PDFF from Fat DESPOT m was lower, with differences of −4.3 ± 4.0 (initial guess from three‐point Dixon) and −1.0 ± 4.0 (initial guess from GC), both statistically significant. Precision was similar between techniques. Standard deviations for PDFF were not significantly different, and significantly higher for R2* (from GC initial guess) only in the 50% nominal fat fraction ROI (F‐test, p=2×10−3). R1f from vials with nominal fat fractions of 25%, 50%, 75%, and 100% using GC initial guesses had significantly higher standard deviations than with Dixon (0.1 vs. 0.11 s−1 on average; p≤0.03). Conversely, the GC approach R1w returned lower standard deviations in vials with nominal fat fractions of 0%, 5%, and 50% (0.033 vs. 0.043 s−1 on average; p≤10−3). Comparing the Models Inspection of the outputs of GC fat‐water separation revealed that the initial phase of fat and water signals were drastically different. This can be seen in examples of initial guess maps from GC and DESPOT1 (Figure S2). This suggests that initial phase should be a distinct free parameter in the complex models and motivated the development of the Fat DESPOT mϕ model. We noted average absolute differences between the ΔB0 maps from FA1 and FA4 in phantoms of up to 3.6 Hz (maximum 26 Hz over all voxels), up to 5.1 Hz in the leg data (maximum 27.3 Hz over all voxels), and up to 7.3 Hz in the liver (maximum 216.6 Hz over all voxels). These maps were therefore entered as a flip‐angle dependent input to the complex Fat DESPOT fits, in Equations (1) and (2). Maps of PDFF, R2*, R1f, and R1w obtained from all model fits to the 8‐echo acquisitions demonstrated high‐quality fits in all ROIs (Figure 4). This being said, the phase sensitive models Fat DESPOT mϕ , Fat DESPOT c and Fat DESPOT cϕ had lower combined nRMSE: 0.08 and 0.14 and 0.14, respectively, compared to Fat DESPOT m with an average value of 0.20. The highest combined nRMSE for a single ROI was 0.33 for Fat DESPOT m , 0.11 for DESPOT mϕ , and 0.20 for both Fat DESPOT c and Fat DESPOT cϕ , corresponding to the tube with a nominal fat fraction of 50%. As with the previously discussed 2 × 6‐echo data fitted with the Fat DESPOT m approach, artifacts appeared in the water compartment of the R2* map regardless of the Fat DESPOT approach. Example model fits using all four models in single voxels of three ROIs are plotted in Figure S3. Multiparametric maps for PDFF, R2*, R1f, R1w, and nRMSE using Fat DESPOT m , Fat DESPOT mϕ , Fat DESPOT c , and Fat DESPOT cϕ on an 8‐echo dataset. To reduce noise in the R1f images, voxels with PDFF < 3% and in the R1w images, voxels PDFF > 97% were masked. Voxels outside the phantom were masked. Parameter estimates from fits to the 8‐echo data showed similar trends versus fat fraction as those from the 12‐echo data. Parameter values are displayed in Figure 5. The most obvious parameter differences were observed for PDFF and R2*, between Fat DESPOT m and the other three models (in ROIs with 25%–75% nominal fat fraction). In select ROIs, ad hoc variations were observed between the models for R1,f and R1,w. Two‐way ANOVA (factors “ROI” and “Model”) for each parameter revealed statistically significant effects for all parameters (p<<10−3). Post hoc testing confirmed that differences between all six pairs of approaches, for all parameters, were statistically significant, with the exception of PDFF and R2* for Fat DESPOT c and Fat DESPOT cϕ . Despite this, PDFF and R2* from phase‐sensitive methods Fat DESPOT mϕ and the complex approaches were far more similar to each other than to those from Fat DESPOT m (Table 1). The only exception was R1f, where Fat DESPOT m and the complex approaches were more similar to each other than to Fat DESPOT mϕ . Box plot of voxel‐wise estimates of PDFF, R2*, R1f, and R1w, using Fat DESPOT m , Fat DESPOT mϕ , Fat DESPOT c , and Fat DESPOT cϕ on an 8‐echo dataset. Box = interquartile range, horizontal line = median, whiskers = data range (=1.5× interquartile range), dots = outliers. TABLE 1Relative difference in estimates of PDFF, R2*, R1f, and R1w between pairs of approaches.ParameterFat DESPOT approach pairings (%) m, mϕ m, c m, cϕ mϕ, c mϕ, cϕ c, cϕ PDFF (%)30.833.033.07.97.90.0 R2* (%)13.414.614.61.61.60.0 R1f (%)16.26.36.114.513.82.7 R1w (%)34.726.025.320.819.92.6 Relative difference in estimates of PDFF, R2*, R1f, and R1w between pairs of approaches. PDFF estimates were significantly different from the reference measurement for all Fat DESPOT models (p<10−18 for all approaches) (Figure 6). Fat DESPOT m exhibited the highest mean error of 3.2% ± 2.3% compared to 1.6% ± 1.3% for Fat DESPOT mϕ and 1.2% ± 1.2% for the complex models. Box plot of the error on the PDFF obtained from Fat DESPOT m , Fat DESPOT mϕ , Fat DESPOT c , and Fat DESPOT cϕ on an 8‐echo dataset from the phantom compared to a reference measurement, as a function of the nominal fat fraction. Box = interquartile range, horizontal line = median, whiskers = data range (=1.5× interquartile range), dots = outliers. Error is significantly different (p < 0.05) between approaches in all ROIs excluding the 75% nominal fat fraction. Overall, Fat DESPOT c resulted in lower standard deviation for PDFF, R1f, and R1w. Fat DESPOT m had the lowest standard deviation for R2*, and Fat DESPOT mϕ had the highest standard deviation for all parameters (Table 2). However, not all ROIs had statistically significant differences in standard deviation and R2* standard deviation differences were only significantly different for 1/7 ROIs when comparing magnitude approaches to the complex approach and 2/7 ROIs when comparing magnitude approaches to each other. Fat DESPOT mϕ had higher variability in standard deviations across ROIs, suggesting lower fit stability. This was particularly noticeable in the precision of R1w estimates for the 50% and 75% tubes. TABLE 2The combined standard deviations for PDFF, R2*, R1f, and R1w across ROIs 1–7 in the variable fat fraction phantom for all four models.ParameterFat DESPOT m Fat DESPOT mϕ Fat DESPOT c Fat DESPOT cϕ PDFF (%)0.140.240.130.13 R2* (s−1)0.440.470.470.47 R1f (s−1)0.280.210.190.19 R1w (s−1)0.0130.0340.0080.09 The combined standard deviations for PDFF, R2*, R1f, and R1w across ROIs 1–7 in the variable fat fraction phantom for all four models. In Vivo Results In vivo Fat DESPOT c and Fat DESPOT cϕ multiparametric maps of the lower leg and upper abdomen (Figure 7) displayed key anatomical features, including the muscle, bone marrow from the tibia, the fibula, the subcutaneous and visceral fat, and the liver, with distinct combinations of Fat DESPOT output values. Initial guess maps for both experiments are presented in Figures S4 and S5. Average Fat DESPOT output parameters from ROIs in each experiment are displayed in Table 3: values from the two complex models were virtually indistinguishable. Multiparametric maps of cross‐sections of the lower leg and of the abdomen for PDFF, R2*, R1f, R1w, and nRMSE using Fat DESPOT c . Voxels outside the leg and the abdomen were masked. TABLE 3PDFF, R2*, R1f, and R1w and combined nRMSE for ROIs in the bone marrow, muscle, and subcutaneous fat of a human lower leg, and in the liver and abdominal subcutaneous fat. Values are presented for both complex models (Fat DESPOT c , Fat DESPOT cϕ ) as mean ± standard deviation.ParameterModelROI# VoxelsBone marrowMuscleLeg fatLiverAbdominal fat143243168147286PDFF (%) c 95 ± 31.9 ± 1.383 ± 114.6 ± 0.994 ± 9 cϕ 95 ± 31.9 ± 1.383 ± 114.6 ± 1193 ± 8 R2* (s−1) c 45 ± 6646 ± 1027 ± 2952 ± 428 ± 9 cϕ 45 ± 6646 ± 1027 ± 2852 ± 428 ± 10 R1f (s−1) c 2.0 ± 0.64.6 ± 3.02.3 ± 0.51.7 ± 0.52.5 ± 0.1 cϕ 2.0 ± 0.64.6 ± 2.92.3 ± 0.51.7 ± 0.52.5 ± 0.1 R1w (s−1) c 1.4 ± 1.70.60 ± 0.030.89 ± 0.51.00 ± 0.072.9 ± 2.1 cϕ 1.3 ± 1.60.6 ± 0.30.89 ± 0.51.03 ± 0.082.6 ± 1.9nRMSEcombined c 0.260.040.240.070.12 cϕ 0.240.040.290.070.11 PDFF, R2*, R1f, and R1w and combined nRMSE for ROIs in the bone marrow, muscle, and subcutaneous fat of a human lower leg, and in the liver and abdominal subcutaneous fat. Values are presented for both complex models (Fat DESPOT c , Fat DESPOT cϕ ) as mean ± standard deviation. Leaving the initial phases as completely free parameters in the Fat DESPOT cϕ resulted in phase differences when compared with the more constrained Fat DESPOT c fit. Mean phase differences of 0.8 and 0.1 rad were observed for fat and water in the phantom, respectively, along with 0.9 and 0.1 rad for fat and water in lower leg, and 0.99 and 0.33 rad in the abdomen. Discussion This study sought to advance signal modeling for multiparametric mapping in fat‐water mixtures. Three signal models were introduced, notably to account for the initial phase of the signal components: one phase‐sensitive magnitude model and two complex models. Fat DESPOT m initialized with GC performed as well or better than with three‐point Dixon. The latter approach requires specific echo‐selection for best performance [36], which imposed a longer 2 × 6‐echo acquisition, while GC is more flexible and therefore is a superior choice for initial parameter estimation in this context. Using the Dixon and GC algorithms to initialize Fat DESPOT m returned similar results with agreement on PDFF and R2* for nearly all ROIs and having small percent differences for R1f and R1w for almost all ROIs. However, the GC and three‐point Dixon approaches disagreed across all parameter in the 50% tube. This suggests that inaccurate initial guesses of PDFF (from three‐point Dixon) or R2* (from a monoexponential fit) may have affected R1f and R1w estimates in this tube. Furthermore, while precision was similar, GC returned significantly lower standard deviations in 2 of 3 parameters. Finally, our GC‐based protocol used a single 8‐echo acquisition scheme rather than the previous 2 × 6 echoes [10]. Reducing the number of acquisitions required for Fat DESPOT increases the clinical feasibility of the approach. The complex approaches, Fat DESPOT c and Fat DESPOT cϕ , which both fit to the complex data starting from initial guesses of PDFF, R2*, B0θ, ϕ0f, and ϕ0w from GC, returned more accurate estimates of PDFF, and more precise estimates of R1f and R1w in phantoms. The improved accuracy of the complex models over the magnitude model in fat‐water separation experiments is consistent with previous findings [47]. The resulting parametric estimates from Fat DESPOT mϕ had a larger standard deviation in some ROIs suggesting they were less stable compared to complex models, or that the complex models benefit from including more data points in the fit. When PDFF values in the phantom were compared to the reference measurement, those from complex approaches were closer than those from the other models. With respect to comparison of R1f and R1w values to reference measurements, our choices were limited: first, MR spectroscopy reveals R1 values that vary across lipid resonances [48], clashing with the assumption of a single R1f in Fat DESPOT models, and second, comparison methods for the work proposed here are also experimental [8, 9, 15, 49, 50] and a reference has yet to be established. The inclusion of phase in the Fat DESPOT model may also be responsible for some of the larger disagreements between approaches. Phase‐sensitive models (Fat DESPOT mϕ , Fat DESPOT c , and Fat DESPOT cϕ ) returned higher R2* compared to Fat DESPOT m in tubes with intermediate nominal fat fraction, while no discernible trend in differences was noted for other parameters. This bias in R2*, paired with visible differences in the initial phase of fat and water from the GC initial guess suggests that error due to the inaccurate modeling of phase may be largely absorbed in the R2* parameter. Furthermore, accounting for phase differences between fat and water resulted in higher quality fits when using both the complex and magnitude data, with Fat DESPOT mϕ having the highest fit quality among the approaches tested. Conversely, the lower precision of Fat DESPOT mϕ compared to all other techniques may be due to insufficient data to compensate for the inclusion of an additional parameter for phase in the model leading to overfitting when using the magnitude signal only. Prior work in our group with a complex model that imposed identical initial phase for water and fat signal components yielded better fits than with zero initial phase but led to inaccurate results with higher excitation pulse FAs [22]; the results from the model presented here with distinct initial phase for water and fat show an improvement in performance, consistent with observations when using excitation FAs greater than five degrees [51]. Comparison of the relaxation rates measured in phantom in this work with the literature is complicated by contradictory trends in prior publications, and the paucity of reports of R1w and R1f. Indeed, some studies found that R1w in gel phantoms was independent of fat fraction [51, 52], consistent with our observations. Others found that both R1f and R1w were fat fraction‐dependent in phantoms [53]. However, the measurement approach and phantom construction, including the use of agar or agarose, dissolved contrast agents, and their respective concentrations, varied between studies, making comparison difficult. Comparison of in vivo measurements with published data was challenging due to limited studies reporting PDFF, R2*, R1f, and R1w in the same anatomy. Nevertheless, our results show generally good agreement with prior work. Comparing tibia bone marrow with femur bone marrow [9], our R1f (2.0 ± 0.6 s−1 for both complex approaches) was lower than the reported value (3.9 ± 0.3 s−1) while the R1w were consistent (1.4 ± 1.7 s−1 for Fat DESPOT c and 1.3 ± 1.6 s−1 for Fat DESPOT cϕ vs. 1.43 ± 0.77 s−1). The uncertainty on R1w in bone marrow was large, weakening the comparison. In subcutaneous fat, R1f (for both complex approaches: 2.3 ± 0.4 s−1 and 2.5 ± 0.1 s−1 in the leg and abdomen, respectively) agreed with reported R1global values—which should be dominated by the fat signal—from one study (2.59 s−1 [24]) but not with another (4.22 s−1 [25]). In muscle, where water content is high, R1w (0.60 ± 0.03 s−1 for both approaches) was close to R1global from one source (0.7 s−1 [54]) but lower than another (1.13 s−1 [26]). In liver, R1w (1.0 ± 0.07 s−1 with Fat DESPOT c and 1.03 ± 0.08 s−1 with Fat DESPOT cϕ ) agreed with one report (1.1 s−1 [55]). For R2* measured in the leg, values in bone marrow (45 ± 60 s−1 for both approaches) and subcutaneous fat (27 ± 29 s−1 for Fat DESPOT c and 27 ± 28 s−1 for Fat DESPOT cϕ ) were broadly consistent with literature (60.8 ± 5.1 s−1 [56] and 23 s−1 [57], respectively), while perhaps more variable. Our R2* in muscle (46 ± 10 s−1 for both approaches) was higher than reported (between 24 and 25 s−1 [26, 58]). In liver, R2* (52 ± 4 s−1) is similar to a prior report (69 ± 22 s−1 [59]). The Fat DESPOT c model uses separate initial phases for water and fat that are held constant for all FAs, while the Fat DESPOT cϕ model uses separate initial phase for each FA, reflecting the dependency of the initial phase on FA and relaxation times [15]. The phase‐sensitive magnitude model features the difference between the two phases of the Fat DESPOT c model. Our goal in contrasting these three models was to weigh the increased accuracy of FA‐dependent phases on accuracy against the risk of unstable parametric estimates with additional free parameters. The two complex models behaved very similarly, in phantom and in vivo. This may be due to the small variations in initial phase observed between FAs with this data. While the difference in the phase of fat between FAs is larger, we believe this is due to the dominance of the water signal in the water compartment of the phantom, muscle, and most of the abdomen; accordingly, the low fat signal is noisy and leads to highly variable phase. Inclusion of FA‐dependent phases may become more important if FA was further increased or R1 values were lower, as this phase difference depends on relaxation times. Interestingly, our choice to accommodate different initial phase angles might be seen to clash with prior findings that enforcing equal initial phase for water and fat is needed for stable fits [60], but that observation was for small FAs (6°). There are some limitations to this study. In our experiments, the fat spectra were not measured directly, which may have affected the accuracy of the PDFF [61]. Our model accounts for a single R2* for fat and water. While including separate R2* values for fat and water has been shown to increase accuracy in some studies [62], it may also lead to less robust fits especially with low‐SNR data [47, 62], which was the case in previous work done by our group (results not shown). Additionally, all approaches to Fat DESPOT appear vulnerable to B0 field inhomogeneity artifacts, notably from the styrofoam insert used in the phantom design. To reduce these issues, alternative initial parameter estimation techniques could be explored. Acquisition time remains a disadvantage in this implementation of Fat DESPOT; however, all acquisitions in this work used at least six signal averages and no parallel imaging, resulting in very high SNR. Reduction of the number of averages and/or introduction of parallel imaging while keeping above this SNR threshold will allow for gains in the acquisition time without reduced fit quality. Furthermore, the number of FAs acquired and used in the fitting algorithm could be reduced [10]. The optimization algorithm was allowed to select relatively high FAs, and this may introduce FA accuracy issues due to slice profile effects, which were not accounted for in this work. The lower leg provided a straightforward site for in vivo measurement, with relatively little motion and a variety of musculoskeletal tissues, while our second in vivo experiment demonstrated the feasibility of abdominal imaging. Further experiments could be conducted in other sites to explore a broader diversity of tissues. Conclusions Phase‐sensitive modeling for Fat DESPOT offered higher accuracy and precision in phantom measurements compared to other versions of this technique and in vivo parametric estimates were comparable to literature. The complex model yielded more accurate PDFF and greater precision for R1w, and R1f than other approaches. To model magnitude data, the inclusion of a term for the initial phase difference between fat and water improves fit quality. The 3D mGRE sequence used for Fat DESPOT is accessible on all clinical scanners, making it highly translatable. The use of the GC algorithm to calculate initial parameter guesses increases the flexibility of echo time selection compared to three‐point Dixon, further simplifying data acquisition for this technique. Hence, the complex approach to Fat DESPOT represents a valuable advancement for multiparametric mapping with potential applications in fatty liver disease, and solid tumor imaging, where measures of R2*, PDFF, R1w, and R1f are of particular value. Funding This work was supported by the Research Institute of the McGill University Health Centre, the Fond de Recherche Québec—Santé (FRQS), and a Discovery Grant from the Natural Sciences and Engineering Research Council of Canada (NSERC). Supporting information Table S1: Sequence parameters for complex and magnitude Fat DESPOT, B1 mapping, and unipolar FW separation in phantom. Table S2: Sequence parameters for complex and magnitude Fat DESPOT, B1 mapping, and unipolar FW separation in vivo. For B1 mapping, TE1 is the effective TE and the # TE is the echo‐train length. Figure S1: A flowchart to illustrate the processing steps of the method. Top: Process for magnitude data, without initial phase term, and initialization using Dixon fat‐water separation. Bottom: Process for phase‐sensitive models using magnitude or complex data, with GC fat‐water separation for initialization. Table S3: Lower and upper bounds for fitting parameters used in Fat DESPOT m , Fat DESPOT c , and Fat DESPOT cϕ . Figure S2: Examples of initial guess and input parameter maps for the variable fat fraction phantom (8‐echo acquisition). The B 1 map was obtained from a dual‐angle method B 1 estimation and R 1global map was obtained from a DESPOT1 algorithm on the upper right. All other estimates were obtained using the Graph Cut algorithm. Four B 0 maps are shown, each of which was obtained by fat‐water separation on data with the various FAs. Figure S3: Examples of voxel‐wise fits for the central pixel of 3 ROIs in the variable fat fraction phantom corresponding to nominal fat fractions of 0% (ROI 1), 50% (ROI 5), and 100% (ROI 7). The left column shows the magnitude of the mGRE data (points) and the Fat DESPOT m fits (dashed line). mGRE data is depicted by points and the Fat DESPOT fits as a dashed lines. Figure S4: Examples of initial guess and input parameter maps for the lower leg (eight‐echo acquisition). PDFF and R2* maps were obtained from Graph Cut fat‐water separation. The R1global map was obtained from a DESPOT1 algorithm. The B 1 map was obtained from a dual‐angle method B 1 estimation. Four B 0 and eight ϕ 0 maps are shown, each of which was obtained by fat‐water separation on data with the various FAs. Figure S5: Examples of initial guess and input parameter maps for the abdomen (six‐echo acquisition). PDFF and R2* maps were obtained from Graph Cut fat‐water separation. The R 1global map was obtained from a DESPOT1 algorithm. The B 1 map was obtained from a dual‐angle method B 1 estimation. Four B 0 and eight ϕ 0 maps are shown, each of which was obtained by fat‐water separation on data with the various FAs.