Laterally Oscillating Trajectory for Undersampling Slices: LOTUS Authors: Mayuri Sothynathan, Paul I. Dubovan, Corey A. Baron Published: 2026 ABSTRACT ABSTRACTPurposeWhile spiral sampling offers SNR advantages for diffusion MRI, its acceleration with simultaneous multislice remains relatively unexplored. This study introduces Laterally Oscillating Trajectory for Undersampling Slices (LOTUS), which is a 3D spiral‐like k‐space trajectory that aims to minimize g‐factor via controlled incoherent aliasing. To aid in validation, we introduce a constrained reconstruction approach that enables robust pseudo‐multiple replica g‐factor estimation for iterative non‐Cartesian reconstructions.MethodsSimulated data sampling of a numerical phantom was performed using LOTUS and several acquisition schemes proposed by others to quantitatively compare the resulting image quality when compared to a known ground truth. Diffusion‐weighted in vivo brain data from two subjects was acquired with two in‐plane acceleration factors (2× and 4×) and two slice acceleration factors (2× and 4×). Estimated g‐factor maps and fractional anisotropy maps were calculated to quantitatively and qualitatively compare trajectory performance. For both simulation and in vivo, reconstructions both with and without compressed sensing were utilized.ResultsSimulations generally showed decreased g‐factor (20%–31%, depending on trajectory, at highest undersampling rate) and improved reconstruction accuracy (mean‐square error, structural similarity index, and entropy metrics) for LOTUS compared to the other trajectories. The in vivo acquisitions demonstrated g‐factor benefits and qualitative image quality improvements that mirrored the simulation results. For both simulation and in vivo, improvements for LOTUS increased for higher numbers of simultaneous slices.ConclusionBy enabling higher rates of slice acceleration, LOTUS shows promise to decrease scan time, which is especially beneficial for diffusion MRI. Introduction Diffusion magnetic resonance imaging (dMRI) is an imaging modality that characterizes the diffusion properties of water in tissue to provide valuable insights on the microstructure of the brain. However, dMRI is plagued by long scan times due to the long repetition time (TR) required to acquire high‐resolution, whole‐brain images for many diffusion directions [1]. Typically, dMRI acquisitions are performed using single‐shot two‐dimensional (2D) Echo Planar Imaging (EPI). Standard 2D parallel imaging approaches are primarily used to reduce distortions due to B0 inhomogeneity and have a relatively limited impact on scan time for dMRI, where a large portion of scan time is spent on diffusion encoding. However, methods to decrease scan times are needed since new dMRI representations and models require many diffusion directions and high diffusion weightings [2]. Simultaneous multislice (SMS) allows for the acquisition of multiple image slices during a single shot, reducing acquisition time by a factor equivalent to the number of simultaneously acquired slices [3]. Additionally, SMS mitigates signal‐to‐noise ratio (SNR) penalties compared to standard parallel imaging approaches since the SNR loss from decreasing the total number of samples is perfectly compensated by the SNR gain from performing a “volumetric” acquisition of multiple slices at once. Thus, the only source of SNR loss is through g‐factor losses, but this can be greatly ameliorated via controlled aliasing approaches that reduce overlap of aliased signal via radiofrequency (RF) or “blipped CAIPI” strategies. Notably, blipped CAIPI is equivalent to 3D sampling of k‐space, where the total span of k‐space acquired in the slice direction is given by the inverse of the spacing between simultaneous slices (Figure 1a) [1, 4]. (a) CAIPI EPI trajectory visualized in 3D, (b) Standard single‐shot spiral trajectory (c) “CAIPI‐like” spiral trajectory for three simultaneous slices, (d) Tilted Hexagonal (T‐Hex) trajectory (e) LOTUS For (c) to (e), SMS rate (R S) = 3. For (e) Cϕ=0.618. Non‐Cartesian spiral sampling (Figure 1b) allows for SNR gains in the order of 40%–80% in dMRI at 3 T due to reduced echo times (TE) resulting from center‐out sampling [5]. Additionally, spiral sampling provides reduced g‐factor compared to EPI [5], approximately isotropic point‐spread functions (PSFs) [6], compatibility with navigator and variable‐density sampling approaches, and increases spatial encoding efficiency which is valuable for extensive signal preparations required for dMRI [7]. Nevertheless, clinical adaptability of spiral imaging is limited due to artifacts from unwanted field perturbations [8]. Field changes caused by B 0 field inhomogeneity may be monitored via B 0 mapping, while recent advancements in field monitoring has enabled the detection of eddy current induced field perturbations [9, 10]. These fields may be accounted for by using an expanded encoding model during image reconstruction to mitigate artifacts [7, 11, 12]. These advancements in field monitoring have increased the feasibility of spiral dMRI and enable the design of new trajectories. Spiral acquisitions with controlled aliasing have previously been explored in the context of 3D volumetric imaging. Distributed spiral trajectories represent a class of 3D k‐space trajectories that traverse a cylinder, sphere, or intermediate shape in k‐space [13, 14], including the 3D spiral staircase trajectory proposed by Anderson et al. [15] Non‐Cartesian blipped CAIPI approaches [3, 4] as well as tilted hexagonal sampling (T‐Hex) [16, 17] have also shown promise to perform SMS with a spiral‐like trajectory. Additionally, preliminary investigations have employed concentric spiral‐like rings in k‐space and the addition of a sinusoidal z‐gradient to a multi‐shot spiral trajectory, which have both shown potential for non‐Cartesian SMS [18, 19]. However, questions remain about the choice of oscillation period for the sinusoidal approach and g‐factor performance for all approaches. This study proposes a 3D k‐space trajectory for SMS inspired by this earlier work [18, 19], but here the period of oscillations is governed in k‐space rather than by the z‐gradient. We compare our proposed technique, Laterally Oscillating Trajectory for Undersampling Slices (LOTUS) (Figure 1e), to CAIPI EPI; standard spiral (i.e., no k z manipulation); “CAIPI‐like” spiral (spiral in k xy and uses gradient blips to sample multiple k z planes, similar to concentric rings and CAIPI EPI); and T‐Hex, a trajectory that is spiral in k xy and samples multiple k z planes using hexagonal grids in planes orthogonal to the read direction [16]. The corresponding gradient zero crossings for each trajectory are shown in Figure S1. To evaluate the performance of the trajectories, we introduce an image reconstruction approach for improving g‐factor computation for non‐Cartesian trajectories that requires no tuning of the number of image reconstruction iterations, in contrast to previous approaches. We find that a LOTUS oscillation period based on the Golden Angle, which evenly distributes k z oscillations throughout the k xy plane, results in minimal noise amplification. Finally, hypothesizing that the pseudo‐random nature of LOTUS may be beneficial for compressed sensing (CS), we assess several image fidelity metrics for compressed sensing reconstructions. Overall, LOTUS has lower g‐factor noise amplification than other trajectories. For low numbers of simultaneous slices, LOTUS and blipped spiral have comparable performance considering both g‐factor and CS performance, but LOTUS exhibits growing advantages over blipped spiral as the number of simultaneous slices increases. Theory Trajectory Design Our implementation of the LOTUS trajectory follows the numerical approach developed by Pipe [13, 20, 21, 22, 23, 24]. The spiral waveform in 2D k‐space (k x, k y) can be denoted in polar coordinates as (1) kxy⇀(t)=kxy(t)eiϕ(t) where kxy⇀ represents the k‐space vector, and kxy and ϕ represents the respective magnitude and phase. For blipped CAIPI, the z‐gradient blips create steps in k‐space separated by Δkz=1/RsΔz, where RS is the number of simultaneously acquired slices and Δz is the slice spacing. Thus, the total kz coverage over all steps is given by Wz=ΔkzRS−1. For LOTUS, we define a sinusoidal oscillation of kz covering this same span of k‐space: (2) kz=Wz2cosCϕϕ where Cϕ is a user‐defined numerical factor that relates the kz oscillation frequency to the spiral rotation frequency. The numerical algorithm developed by Pipe creates a discretized trajectory from start to finish one point at a time. For each point, the key equations governing the gradient values are [25]: (3) Gxy→(t)=Gxy(t)eiθ(t) (4) dkxydϕ=Δkxy2π (5) dϕdt=γGxysin(α)kxy (6) dkxydt=γGxycos(α) where Gxy and θ are the magnitude and angle of the gradient in the xy‐plane, Δkxy is the spacing between successive spiral loops (governed by Nyquist criteria and in‐plane undersampling rate), and α=θ−ϕ is the angle of the gradient relative to the angle of k⇀(t). Taking the derivative of Equation (2) with respect to ϕ and multiplying the result with Equation (5) yields: (7) dkzdt=WzγGxysin(α)CϕsinCϕϕ2kxy Recognizing that dkz/dt=γGz, we arrive at the value of Gz relative to the in‐plane spiral parameters: (8) Gz(t)=Gxy(t)CϕWz2kxy(t)sin(α(t))sinCϕϕ(t) Accordingly, to generate a LOTUS trajectory, Equation (8) is used to find G z for every step in the numerical algorithm (immediately after G x(t) and G y(t) are determined for that step), and constraints for slew rate and maximum gradient consider the net values over all three axes. Note that it is important to define kz based on cosine rather than sine so that gradient values are close to zero at the beginning of the trajectory where t=0. However, without intervention this would cause kz to oscillate between 0 and Wz, and thus to center kz about 0 we apply an additional factor of 0.5 to Gz(t) for 0≤Cϕϕ(t)<π. G‐Factor Estimation for Non‐Cartesian MRI Expanded encoding reconstructions of non‐Cartesian data ostensibly solve the equation (9) x^=argminx‖SCx−y‖22 where x is the object‐domain image, y is the k‐space data, S represents data sampling including B0 inhomogeneity, and C represents receiver sensitivity profiles [7, 26]. Iterative reconstructions using the conjugate gradient method are typically used for these problems, but it is a well‐known issue that, without regularization, noise will overwhelm the image for single‐shot spiral acquisitions. Accordingly, regularization is performed to control noise, either implicitly through early stopping of conjugate gradient iterations [27, 28] or other explicit types of regularization [29]. From our observations, noise is exacerbated by the algorithm attempting to fill in the outer unacquired corners of k‐space. Since these regions are not directly sampled, their estimation is poorly conditioned under the combined sampling operator and receiver sensitivity profiles (Figure S2). Notably, the noise propagates into the acquired region of k‐space as iterations are increased and simply masking the corners of k‐space after reconstruction fails to fully compensate for it. These issues present a problem when attempting to characterize g‐factor, as the number of iterations strongly affects g‐factors and can even lead to g<1 when too low (Figure S3). Here, we propose a simple modification to the above equation to explicitly restrict the reconstruction to only fill in the circular k‐space region that has been acquired and avoid extraneous noise amplification: (10) x^=argminx‖SCx−y‖22+λkMkFx22 where F is a 2D Fourier transform to bring the estimated image to Cartesian k‐space, Mk is a mask that selects only the unacquired corners of k‐space, and λk is the strength of suppression of the unacquired parts of k‐space. We chose λk as a very large value that is effectively infinite to completely suppress the unacquired portions of k‐space, and the conjugate gradient method was permitted to fully converge. It should be noted that this reconstruction constraint is not intended as a general‐purpose denoising regularizer (i.e., Tikhonov or sparsity‐based regularization), but rather to prevent ill‐conditioned extrapolation of signal into unsupported k‐space regions and thereby stabilize subsequent g‐factor analysis (Figure S3). To assess g‐factor related noise amplification, the pseudo‐ multiple replica method was used, where each replica was created by adding a new random realization of noise to the data before evaluating Equation (10) [30]. Noise standard deviation maps (σ) were computed and normalized by the square root of the readout duration to account for differences in acceleration between trajectories and obtain the “normalized noise map,” NNM: (11) NNM=σ/TRO where TRO is readout time. Note that normalization by TRO is a generalization of the traditional normalization by the square root of the acceleration rate, which is required for non‐Cartesian MRI where the acceleration rate is not necessarily an integer value. For simulations, g‐factor maps were computed via (12) g=NNM/NNMF where NNMF is the normalized noise map for a fully sampled reference scan. For in vivo scans, a fully sampled reference scan had a very long readout time leading to excessive T2* decay (∼90 ms). Accordingly, here NNM was normalized by the mean NNM (averaged over the entire brain in all slices) computed from NNMF^, a scan with no SMS and an in‐plane undersampling rate of 2, which is expected to have low g‐factor noise amplification, to obtain an approximation of g‐factor, g^: (13) g^=NNMNNMF^ Compressed Sensing Reconstruction While pure random sampling is impractical due to hardware and physiological constraints, it is expected that LOTUS will have an incoherent PSF similar to random sampling. It is hypothesized that this will increase the efficacy of CS, a reconstruction algorithm that increases acquisition efficiency by enabling k‐space undersampling. CS solves: (14) x^=argminx12‖SCx−y‖22+λ‖ψx‖11 where ψ is a wavelet transform, and λ tunes the CS strength [31]. Notably, the regularization used for CS implicitly suppresses noise from the unacquired corners of k‐space, and we have found additional regularization with λkMkFx22 has a negligible effect on the reconstructed image. Methods Trajectories CAIPI‐EPI acquires multiple planes of k‐space using gradient blips (Figure 1a). Similar to CAIPI‐EPI, A “CAIPI‐like” spiral trajectory samples multiple planes of k‐space [3, 4] (Figure 1c) and T‐Hex [16, 17] (Figure 1d) samples multiple planes of k‐space using a hexagonal grid in planes orthogonal to the read direction (Figure S1b). Meanwhile, LOTUS (Figure 1e) has sinusoidal k z with zero‐crossings dictated by the period factor Cϕ (Figure S1c). In experiments, the in‐plane acceleration rate, RIP, is defined from the separation between successive rings of the spiral, ignoring kz. The slice‐acceleration rate, Rs, is the number of simultaneous slices. Reconstruction Image reconstruction for both simulations and in vivo scans (details below) was performed in MATLAB using the open‐source MatMRI toolbox, using an iterative expanded encoding model‐based reconstruction [29, 32, 33, 34]. MatMRI is a graphics processing unit (GPU)‐enabled reconstruction framework that can handle any k‐space trajectory with the option to include a B 0 map, higher‐order coefficients fitted from the monitored field dynamics, and implement CS (ℓ 1‐norm regularization). CS image reconstructions of in vivo data used an automatically tuned regularization factor [29, 31] while a range of regularization factors were explored for simulations because a ground truth image was available for comparison. The delay between the field probe and MRI data was determined using a model‐based retrospective algorithm [35]. All experiments were run on a workstation with an Intel i9‐11900K processor and 128 GB random‐access memory with GPU Nvidia GeForce RTX 4090 with 25.76 GB of GDDR6X memory. Simulations To simulate reconstructions with a known ground truth, a numerical phantom was generated in MATLAB and used as x in the forward encoding model =SCx, to simulate k‐space data. C was created from 16 synthetic receivers with sensitivity decreasing with an inverse‐square trend relative to a point approximately a factor of 1.5 outside the FOV. To generate noisy data, complex Gaussian noise was added to the simulated data (SNR for fully sampled spiral ∼33). Simulations were performed for spiral, CAIPI spiral, T‐Hex, and LOTUS acquisitions. A CS reconstruction was performed using MatMRI; however, they did not include a B 0 map and second‐order or higher coefficients of the expanded encoding model. Investigation of k z Oscillation Period To determine the ideal k z oscillation period of LOTUS, g‐factor maps and mean g‐factor were calculated using reconstructions that employ Equation (10) for R S = 5 and RIP = 2. Cϕ from 0 to 1 in 0.1 increments as well as Cϕ = 0.618 (inverse of Golden ratio) were used. To quantitatively evaluate quality between CS reconstructions that employed Equation (14), Mean Square Error (MSE), Structural Similarity Index (SSIM) and Entropy versus λ were plotted for each candidate oscillation period Cϕ. λ was varied from 0.01 to 0.063. Based on these results, Cϕ resulting in greatest reduction in error was used for subsequent comparison against other trajectories. Trajectory Comparisons Point Spread Functions To shed light on the controlled aliasing introduced by each trajectory, PSF were constructed via PSF = SH1 where 1 represents an input vector of ones and H signifies a conjugate transpose [6] and convolved with numerical phantoms for spiral, CAIPI spiral, and LOTUS acquisitions for RS = 5. G‐Factor Maps and Error Metrics G‐factor was calculated for RIP = 1, 2, and 3 and Rs = 2 and 5. To quantitatively evaluate quality between CS reconstructions and the ground truth, MSE, SSIM, and entropy were calculated for λ varying from 0.010 to 0.063. Both 2 and 5 simultaneous slices were investigated, with RIP = 2. The remaining parameters for all simulations were: FOV = 200 × 200 mm2, in‐plane resolution = 1.5 × 1.5 mm2, maximum gradient strength = 30 mT/m, and maximum slew rate = 120 mT/m/msec. In Vivo Two healthy volunteers were scanned on a 3 T MRI scanner (Siemens Magnetom Prisma) at Western University's Centre for Functional and Metabolic Mapping (80 mT/m maximum gradient strength and 200 T/m/s maximum slew rate). This study was approved by the Institutional Review Board at Western University, and informed consent was obtained prior to scanning. Field monitoring was performed using a 19F commercial field probe system with 16 probes mounted on a scaffold that optimizes placement (Skope Clip‐on Camera) to obtain field dynamics that were fit to second‐order spherical harmonics [11]. The field monitoring scans were performed separately from subject scanning using identical sequence parameters. The collected k‐space data were coil compressed to 20 virtual coils to improve reconstruction speed [36, 37, 38, 39], and noise correlation between receivers was corrected using pre‐whitening before reconstructions [40]. A Cartesian dual‐echo gradient‐echo acquisition was used to estimate B 0 maps for inclusion in S in a model‐based CS reconstruction to correct for static off‐resonance effects. The imaging parameters were as follows: FOV = 240 × 240 mm2, spatial resolution = 1.5 mm isotropic, and TE1/TE2 = 4.08/5.10 ms. Sensitivity coil maps were estimated from the first echo using ESPIRIT [41]. To demonstrate the impact of B 0 correction and higher‐order field monitoring, additional reconstructions were performed without B 0 maps and without higher‐order field probe corrections for spiral acquisitions with R S = 1 and R IP = 2 (Figure S4). Diffusion‐weighted acquisitions were performed, comparing CAIPI EPI (vendor diffusion MRI sequence), CAIPI Spiral (in‐house developed, starting from vendor CAIPI EPI), and LOTUS (in‐house developed, starting from vendor CAIPI EPI) trajectories for: (a) RS = 4, RIP = 2, TR = 7500 ms, scan time = 1 min; and (b) RS = 2, RIP= 4, TR = 15 000 ms, scan time = 2 min. TR was doubled for RS = 2 to keep interslice crosstalk comparable between the two cases. The readout times for CAIPI EPI, CAIPI Spiral, and LOTUS for RS = 4 and RIP = 2 were 43, 47, and 46 ms, respectively. The readout times for CAIPI EPI, CAIPI Spiral, and LOTUS for RS = 2 and RIP = 4 were 14, 19, and 19 ms, respectively. The corresponding gradient trajectories are shown in Figure S5. One 2D spiral acquisition was conducted as a reference scan with RS = 1, RIP = 2, TR = 7500 ms, and readout time = 47 ms. The remaining common parameters were: FOV = 218 × 218 mm, 1.7 mm isotropic resolution, 60 slices, and TE = 62 ms. The imaging protocol used monopolar pulsed gradient spin‐echo encoding using a b‐value of b = 1000 s/mm2 and 6 diffusion directions, plus one b = 0 s/mm2. Separate image reconstruction that employed either Equation (10) (i.e., k‐space masking) or Equation (14) (i.e., CS) were performed using the MatMRI toolbox [29, 32]. Mean diffusion‐weighted images were computed by taking the average over the six diffusion encoding directions. Following the reconstruction, g^ was computed using the pseudo‐multiple replica method with reconstructions that employ Equation (10) [30]. All image reconstructions were allowed to fully converge (i.e., no early stopping implicit regularization [28] was performed). Mean (g^mean) and peak (g^peak) g‐factors were computed in a large region‐of‐interest manually selected to contain all the brain tissue across all 60 slices. The peak value represents the 95th percentile, indicating that 95% of the g‐factor values fall below this threshold. Only the b = 0 scans were used for g‐factor calculations. The MRtrix3 [42] package was used to estimate the diffusion tensor and obtain FA. Results Simulations Investigation of k z Oscillation Period Figure 2a shows the LOTUS trajectories for Cϕ of 0.1, 0.3, 0.618, and 0.9, (b) shows mean g‐factor for all simulated Cϕ values, and (c) shows the g‐factor maps for Cϕ of 0.1, 0.3, 0.618, and 0.9. Significant improvements in g‐factor are observed for Cϕ= 0.3 and 0.618. Figure 2d shows MSE, SSIM, and Entropy plotted against λ for each of the oscillation periods depicted in Figure 2a. Overall, maximum reduction in error is observed in all metrics for Cϕ= 0.618, which is the inverse of the golden ratio. Thus, this value was used when comparing LOTUS against other trajectories. (a) LOTUS trajectory visualized in 3D for (i) CΦ=0.1, (ii) CΦ=0.3, (iii) CΦ=0.618, (iv) CΦ=0.9 (b) Mean g‐factor vs. CΦ for CΦ values ranging from 0.1 to 0.9 in increments of 0.1 (c) g‐factor maps of a sample slice for (i) CΦ=0.1, (ii) CΦ=0.3, (iii) CΦ=0.618, (iv) CΦ=0.9 (d) Simulations of compressed sensing reconstructions of a numerical phantom, showing (i) Mean Square Error (MSE), (ii) Structural Similarity Index (SSIM), and (iii) Entropy vs. CS tuning parameter λ for CΦ=0.1,0.30.618,and0.9. Constant parameters include SMS rate RS = 5 and in‐plane undersampling rate RIP= 2. Comparison of Trajectories Point Spread Functions Figure 3a shows the PSF for each simultaneously acquired slice for spiral, CAIPI Spiral, T‐Hex, and LOTUS for RS = 5. PSFs show more incoherent aliasing for LOTUS compared to the other trajectories. Increased incoherent aliasing is observed in outer slices for all trajectories but spiral due to the larger field offsets from the z‐gradients applied for controlled aliasing. Figure 3b shows the PSF from each slice in Figure 3a convolved with a numerical phantom. Note that the image acquired from simple gridding of the data with no parallel imaging reconstruction is the sum of the images in Figure 3a across all slices, and controlled aliasing causes overlapping signal to be distributed through the FOV. (a) Point spread functions (PSF) comparing spiral, CAIPI spiral, and LOTUS acquisition for SMS rate RS = 5 (each column is a different simultaneously acquired slice); (b) Corresponding PSFs convolved with numerical phantom. Note that different slices were created by applying affine transformations to the standard Shepp‐Logan phantom. G‐Factor Maps Figure 4 shows g‐Factor maps for RS = 2 and RS = 5. The center slice is shown as a sample for R IP = 1, 2, and 3. Mean g‐factor (over all slices) is shown below each computed image. G‐factors were comparable for RS = 2, but significant improvement is seen using 3D non‐Cartesian trajectories for RS = 5. LOTUS exhibits dramatically reduced g‐factors for RS = 5. g‐factor maps for a sample slice of a numerical phantom for Spiral, CAIPI Spiral, and LOTUS trajectories. In‐plane undersampling rates (R IP) of 1, 2, and 3 for each SMS rate (R S) (a) RS = 2 and (b) RS = 5 were investigated. Large improvements for LOTUS over the other trajectories are observed with many simultaneous slices. Mean g‐factor (gmean) and peak g‐factor (gpeak) across all simultaneous slices are reported for each case. Compressed Sensing Image Reconstructions Figure 5 compares spiral, CAIPI Spiral, T‐Hex, and LOTUS trajectory performance quantitatively using MSE, SSIM, and entropy for both RS = 2 and RS = 5, both with R IP = 2. All error metrics showed comparable results for RS = 2 between trajectories. However, LOTUS performed best in all metrics with increased slice acceleration. Figure 6 compares qualitative performance for R S = 5. Coherent artifacts and blurring near edges in the phantom can be observed, which are mostly benign with LOTUS (Figure 6b). Simulations of compressed sensing reconstructions of a numerical phantom, showing (a) Mean Square Error (MSE), (b) Structural Similarity Index (SSIM), and (c) Entropy vs. CS tuning parameter λ for SMS rate RS = 2 and RS = 5 and an in‐plane undersampling rate of 2. (a) Simulations of compressed sensing reconstructions of a numerical using λ value that minimizes MSE for RS = 5 (b) Corresponding difference images showing areas with more artifact (red arrows) and areas where image features are better preserved (blue arrows). In Vivo Figure 7 shows mean diffusion‐weighted images and corresponding approximated g‐factor maps for each candidate trajectory for RS = 2 and R IP = 4, while Figure 8 shows the corresponding results for RS = 4 and R IP = 2. The diffusion‐weighted images were reconstructed with k‐space masking Equation (10) and averaged over the six diffusion encoding directions. The corresponding approximated g‐factor (g^) maps were computed using the pseudo‐multiple replica method with reconstructions based on Equation (10). The qualitative differences between trajectories were subtle in the diffusion‐weighted images. The non‐Cartesian trajectories look slightly blurred compared to EPI due to the broader T2* PSF associated with spiral [6]. Meanwhile, both 3D non‐Cartesian trajectories, CAIPI Spiral and LOTUS, exhibit large g‐factor advantages compared to CAIPI EPI. It is difficult to discern differences in performance between CAIPI spiral and LOTUS for RS = 2 and R IP = 4 in Figure 7; however, g‐factor advantages were evident for LOTUS for RS = 4 in Figure 8, with a mean value 11% lower than CAIPI spiral. These observations suggesting that the benefits of LOTUS are less evident with increased in‐plane acceleration, but more evident with increased slice acceleration, were similar to the simulations. Similar results were observed in both subjects. Diffusion‐weighted images (Equation 10) of simultaneously acquired slices and corresponding approximated g‐factor maps (Equation 13) for prospectively acquired in vivo scans for SMS rate RS = 2 and in‐plane Undersampling rate (R IP) = 4. For each acquisition, g^mean and g^peak were calculated across all 60 slices. Diffusion‐weighted images (Equation 10) of simultaneously acquired slices and corresponding approximated g‐factor maps (Equation 13) for prospectively acquired in vivo scans for SMS rate RS= 4 and in‐plane Undersampling rate (R IP) = 2. For each acquisition, g^mean and g^peak were calculated across all 60 slices. Figure 9 shows the FA maps color coded by eigenvalue direction for images computed using Equation (10) for EPI CAIPI, CAIPI Spiral, and LOTUS acquisitions. Comparison of the FA maps revealed that LOTUS acquisitions demonstrate reduced noise in comparison to the EPI CAIPI and CAIPI spiral acquisitions. These trends were especially visible for Subject 1 whose brain filled more of the FOV, which is expected because smaller objects will have smaller g‐factor noise amplification due to more effective controlled aliasing. Benefits were more evident with increased slice acceleration. Standard FA maps are provided in Figure S6. Sample slice of FA Maps color coded by eigenvector direction determined using Equation (10) for two subjects. Images were acquired using EPI CAIPI, CAIPI Spiral, and LOTUS for (a) R S = 2 and R IP = 4 and (b) R S = 4 and R IP = 2. Figure 10 shows the FA maps color coded by eigenvalue direction for the images that were computed using CS reconstructions for EPI CAIPI, CAIPI Spiral, and LOTUS acquisitions. Comparison of the FA maps revealed that LOTUS acquisitions demonstrate reduced artifact in comparison to the EPI CAIPI and CAIPI spiral acquisitions. Again, benefits were more evident with increased slice acceleration and larger brain size. Standard FA maps are provided in Figure S7. Sample slice of FA Maps color coded by eigenvalue direction (determined using a compressed sensing reconstruction employing Equation (14)) for two subjects. Images were acquired using EPI CAIPI, CAIPI Spiral, and LOTUS for (a) SMS rate (R S) = 2 and in‐plane undersampling rate (R IP) = 4 and (b) R S = 4 and R IP = 2. Discussion Simulated and in vivo results both indicate benefits of LOTUS for high rates of slice acceleration, which can likely be attributed to the controlled aliasing varying more between slices. However, for low rates of slice acceleration all trajectories had comparable performance. MSE was marginally larger for LOTUS in simulated results for RS = 2, but it had comparable SSIM to the other SMS trajectories. However, all differences for RS = 2 were small, suggesting that for low rates of slice acceleration, trajectory dependent differences are minimal. Additionally, at high in‐plane acceleration (R IP = 3) and low rates of slice acceleration (R S = 2), the g‐factor was lowest for standard spiral (Figure 4). This surprising result suggests that the addition of controlled aliasing slightly impairs the ability to account for a large in‐plane acceleration, possibly because it causes larger gaps between samples for any given k z plane. That said, the differences between the trajectories for low rates of slice acceleration were subtle. Incoherence in the sampling pattern, or “noise‐like” artifact in the sparsifying transform domain, is also essential for CS efficacy. Theoretically, CS would be most effective with a completely random subset of k‐space, providing low coherence. However, since this is not practical due to hardware and physiological constraints, sampling trajectories are designed to follow relatively smooth lines and curves. 2D EPI acquisitions are limited because they require full sampling in the readout direction. Thus, the design of incoherent sampling patterns is limited to only one degree of freedom, along the phase encoding direction. In 2D non‐Cartesian acquisition, undersampling is distributed across x and y which increases incoherence, and in 3D non‐Cartesian acquisitions, there are 3 degrees of freedom to design incoherent sampling patterns. LOTUS performed better than CAIPI spiral in CS reconstructions, which could be due to greater incoherence enforced in 3D. However, it is generally difficult to discern if improved CS performance is due to this incoherence or the reduction in g‐factor, since these two sampling features are strongly linked. That said, the changes with Cϕ provide evidence that the pseudo‐random sampling provides a benefit for CS beyond g‐factor, given that even though g‐factor is comparable for Cϕ of 0.3 and 0.618, the value of 0.618 provides slightly lower MSE in CS reconstructions (Figure 2). This can likely be attributed to the increased incoherence in the z‐direction. To investigate this hypothesis further, PSFs were computed for LOTUS for Cϕof0.3and0.618. These were compared to other candidate trajectories by computing their entropy (Figure S8). Pure random sampling would result in a PSF with maximal entropy, and increasing coherence (e.g., via high amplitude rings observable in many of the PSFs in Figure 3) will decrease the entropy. Here, highest entropy is observed for LOTUS with Cϕ=0.618, thus supporting the supposition that a value of Cϕ based on the golden ratio maximizes incoherence, which improves g‐factor and may modestly improve CS performance. Much larger g‐factor noise amplification is observed in the in vivo EPI acquisitions, which is expected because in‐plane undersampling only occurs in one direction compared to the undersampling being distributed across x and y for spiral [5]. The distribution of undersampling in 3D further improves noise amplification for CAIPI Spiral and LOTUS. It should be noted that the large noise amplification was not obviously evident in the diffusion‐weighted images because a high SNR acquisition with large slice thickness was used to maximize the quality of the g‐factor computations, which was the primary concern for the experiments. Nevertheless, qualitative differences were still evident in FA even at this high SNR. Thus, although the reported ∼0.3 improvement in g‐factor at R S = 4 produces modest visual differences in diffusion‐weighted images, it remains relevant for subsequent diffusion‐tensor metrics, such as FA, which are inherently more sensitive to noise and artifacts. LOTUS has been formulated synergistically with the spiral gradient equations and requires no other parameters to be set other than Cϕ, which our results suggest can be fixed to a value of 0.618. Blipped approaches, on the other hand, require setting blip durations and slew rates. We have found practical challenges with blipped approaches, because blips overlap near the center of the spiral if blips are too long in duration (resulting in trajectory generation failure), but readout duration is lengthened if too many gradient resources are devoted to them to make them shorter (i.e., devoting net slew to blips reduces the slew available to turn the spiral in‐plane). As a result, we needed to tune blip parameters for each trajectory, which was not required for LOTUS. It is also worth noting that LOTUS is also compatible with alternative spiral‐like approaches, such as variable density spiral, requiring no changes from the formulation introduced here. Finally, we have also found that LOTUS generally has very little impact on readout duration. For example, at a slice acceleration factor of R S = 5, the readout duration was 8.57 ms for LOTUS and 8.54 ms for the conventional spiral, a negligible difference of just 0.03 ms. Conclusion In this work, we introduced a 3D non‐Cartesian sampling trajectory, referred to as “Laterally Oscillating Trajectory for Undersampling of Slices” (LOTUS). Trajectory performance was compared to other trajectories in diffusion MRI at 3 T, and g‐factors were assessed. Simulations and in vivo acquisitions indicated benefits of LOTUS compared to blipped approaches, particularly for high rates of slice acceleration. By enabling higher rates of slice acceleration, LOTUS shows promise for decreasing scan time, which is especially beneficial for diffusion MRI. Funding This work was supported by the Ontario Graduate Scholarship, NSERC Discovery Grants (RGPIN‐2025‐05597). Supporting information Figure S1: (a) “CAIPI‐like” Spiral visualized in x‐y plane with k z zero‐crossings indicated in red (b) T‐Hex k x zero crossing visualized in y‐z plane (c) LOTUS visualized in x‐y plane with k z zero‐crossings indicated in red. For (a) to (c), simultaneous multislice rate (R S) = 3. For (c), Cϕ=0.618. Figure S2: From the top row, the proposed method to explicitly enforce no k‐space signal in unacquired corner regions (Equation 10) enables full convergence of conjugate gradient iterations without excessive noise amplification. Rows 2–5 show that with the standard approach (Equation 9), noise amplification increases with increasing numbers of conjugate gradient iterations. The right half of k‐space and object‐domain images show the difference from the top row (scaled by 2×), and σ is the noise standard deviation determined using the pseudo‐multiple replica method. Notably, masking the corners of k‐space after finishing iterations (Columns 4–6) does not fully account for these noise increases at high numbers of iterations, which also occur within the circular region of acquired k‐space. Even though the noise increases are most severe in the corners of k‐space, the noise increases within the acquired region of k‐space may occur because the algorithm will inherently seek a solution without sharp edges in k‐space (since this would cause rippling in the object domain that does not agree with the smooth receiver profiles and true underlying image). Figure S3: Comparison of reconstruction strategies for spiral MRI under different in‐plane undersampling rates. Top rows show reconstructed magnitude images, while bottom rows show corresponding approximated g‐factor (g^) maps estimated using the pseudo‐multiple replica method. Results are shown for k‐space masking (regularization term λkMkFx22, Equation 10), early stopping of conjugate gradient iterations using Equation (9) (5, 30, 150 iterations), and Tikhonov regularization (regularization term λT‖x‖22) with varying regularization strengths (λT = 10−4, 10−2, 1). Early stopping and Tikhonov regularization both require tuning of a parameter (number of iterations and regularization strength, respectively) that affects both the image reconstruction and g‐factor. Improper tuning to too small or large parameter values is observed to reduce g‐factor. Explicit k‐space masking requires no tuning beyond ensuring that the regularization factor is large enough (we found no changes to images and maps for increasing λk beyond ∼1 and used λk = 10), mitigating variability of g‐factor due to user‐defined parameters. Figure S4: Comparison of spiral reconstructions with and without correction for off‐resonance and trajectory errors. From left to right: fully corrected reconstruction (including B 0 map and higher‐order field probe correction), reconstruction without higher‐order field probe correction, and reconstruction without B 0 map correction (with higher‐order terms). Figure S5: Gradient trajectories for EPI CAIPI, Spiral CAIPI, and LOTUS scans conducted for (a) simultaneous multislice rate (R S) = 2 and in‐plane undersampling rate (R IP) = 4 and (b) R S = 4 and R IP = 2, measured using field monitoring. Figure S6: Standard FA maps corresponding to those shown in Figure 9. Figure S7: Standard FA maps corresponding to those shown in Figure 10. Figure S8: (a) Point spread functions (PSF) comparing LOTUS acquisition where simultaneous multislice rate (RS) = 5 for (i) CΦ=0.3 and (ii) CΦ=0.618, and (b) histogram comparing entropy for PSFs comparing spiral, CAIPI spiral, T‐Hex and two LOTUS acquisitions (CΦ=0.3and0.618) for RS = 5.