Variable-Density Spiral Sampling in MRI

• Variable-density spiral sampling is a design paradigm that adjusts k-space density by oversampling low-frequency regions to improve compressed sensing recovery and minimize artifacts.
• It employs power-law and piecewise density functions with analytic coordinate discretization to meet MRI hardware constraints and optimize gradient waveforms.
• Integration with deep learning reconstruction techniques has shown enhanced image quality (e.g., SSIM > 0.87) under high acceleration, despite optimization challenges.

Variable-density spiral sampling is a design paradigm for continuous, non-Cartesian sampling trajectories, particularly in magnetic resonance imaging (MRI), that modulates the spatial distribution of samples in k-space to efficiently balance image fidelity, scan acceleration, and the physical constraints of MRI hardware. By localizing denser sampling to low spatial frequencies and sparser coverage at the periphery, variable-density spirals enable both improved compressed sensing recovery and artifact suppression, ultimately facilitating high temporal and spatial resolution imaging even under extreme undersampling regimes. The technique integrates theoretical sampling statistics, hardware-limited trajectory synthesis, and data-driven or analytic optimization to realize practical sequences for 2D and 3D MRI, either as isolated acquisition strategies or in conjunction with advanced reconstruction algorithms.

1. Mathematical Formulation and Theory of Variable-Density Spiral Trajectories

A general spiral trajectory in complex k-space is parameterized as $k(t) = r(t)\, e^{i\theta(t)}$, where $r(t)$ is the radial profile and $\theta(t)$ the azimuthal angle. Variable-density spirals are constructed so that the local sampling density falls off with radius, typically by setting $r(t) = \lambda t^{\alpha}$, resulting in a sample density proportional to $\rho(r) \propto r^{1/\alpha - 1}$—thereby oversampling low |k| (center) regions for $\alpha > 1$ (Chan et al., 2024).

Alternative density laws can be imposed, such as

$$\rho(r) \propto \left[1+\left(\frac{r}{R}\right)^p\right]^{-1},$$

where $R$ is the maximum k-space radius and $p$ controls the fall-off sharpness (Chan et al., 2024). More piecewise constructions, as in the HyperSLICE framework, split k-space into low-frequency (radius $\rho \leq \rho_{in}$) and high-frequency ($r(t)$0) regions, with a smooth transition between them, using explicit acceleration factors $r(t)$1, $r(t)$2 (Jaubert et al., 2023).

In 3D, analytic coordinate discretization generates a continuous spiral on, e.g., a unit cylinder ($r(t)$3), and discretizes it into a set of interleaved planar spirals—each defined by ($r(t)$4)—ensuring prescribed sampling densities in both in-plane and through-plane directions (Jang et al., 21 Mar 2025).

2. Density Optimization and Discretization Algorithms

Selecting optimal density profiles requires trading off low-frequency fidelity with high-frequency resolution. The power-law exponent $r(t)$5 or the piecewise parameters ($r(t)$6) are typically optimized empirically. In (Chan et al., 2024), a grid search over the number of interleaves $r(t)$7 and the exponent $r(t)$8 maximized the structural similarity index (SSIM) in reconstructions, yielding best results for $r(t)$9 and L ≈ 23 for a 20 ms acquisition, with SSIM > 0.87. The HyperSLICE approach employed HyperBand (Jaubert et al., 2023) to jointly optimize six hyperparameters of the spiral density and temporal structure, with the best configuration involving $\theta(t)$0 and Hanning transition.

Discretization of the continuous spiral is performed by selecting points $\theta(t)$1, evaluating ($\theta(t)$2), and associating to each a rotated template spiral. The total number of readouts $\theta(t)$3 is given by integrating the plane- and radius-dependent sampling requirements, e.g.,

$\theta(t)$4

where $\theta(t)$5 is the number of interleaves per transverse plane, itself determined by Nyquist sampling and the chosen density law (Jang et al., 21 Mar 2025).

3. Theoretical Sampling Guarantees and Sampling Subspace Construction

Variable-density spiral sampling can be recast within the general theory of variable density samplers (VDS) (Chauffert et al., 2013). For a stochastic process $\theta(t)$6 generating k-space locations, the empirical measure of visited points converges to the density $\theta(t)$7, ensuring prescribed spatial statistics. In compressed sensing, mixed deterministic/random strategies that deterministically select high-coherence (low-frequency) measurements and randomly sample elsewhere—according to the theoretically optimal $\theta(t)$8—break the coherence barrier, reducing the number of measurements needed for exact recovery.

Continuous VDS strategies include:

• Random-walk-based samplers, which have provable but slow-mixing convergence to $\theta(t)$9.
• TSP-based trajectories, where a curve solving the Traveling Salesman Problem (TSP) on samples drawn from $r(t) = \lambda t^{\alpha}$0 achieves the target density in the limit.

For spiral sampling, the empirical density converges to target $r(t) = \lambda t^{\alpha}$1 when the radial profile $r(t) = \lambda t^{\alpha}$2 is chosen such that

$r(t) = \lambda t^{\alpha}$3

enforced over the trajectory dwell time (Chauffert et al., 2013). Polynomial decay profiles, e.g., $r(t) = \lambda t^{\alpha}$4 in 2D, are repeatedly found near-optimal for compressed sensing MRI.

4. Hardware Constraints, Gradient Waveform Synthesis, and Practical Implementation

The MRI system’s gradient amplitude ($r(t) = \lambda t^{\alpha}$5) and slew rate ($r(t) = \lambda t^{\alpha}$6) are hard physical constraints on achievable k-space velocity and acceleration:

• $r(t) = \lambda t^{\alpha}$7
• $r(t) = \lambda t^{\alpha}$8 with $r(t) = \lambda t^{\alpha}$9, $\rho(r) \propto r^{1/\alpha - 1}$0 gyromagnetic ratio (Chan et al., 2024, Jang et al., 21 Mar 2025).

The desired spiral parameterization must be resampled in time to ensure that the gradient and slew profile of the synthesized waveform does not violate these limits. Practical algorithms iteratively resample the spatial spiral—by either scaling time directly or optimizing via parametric methods—until the output waveform $\rho(r) \propto r^{1/\alpha - 1}$1 is hardware-compatible (Chan et al., 2024, Jaubert et al., 2023).

In the analytic coordinate discretization approach, these constraints are satisfied by precomputing the in-plane k-space extent $\rho(r) \propto r^{1/\alpha - 1}$2, desired FOV scaling laws ($\rho(r) \propto r^{1/\alpha - 1}$3), and solving ordinary differential equations for $\rho(r) \propto r^{1/\alpha - 1}$4, $\rho(r) \propto r^{1/\alpha - 1}$5 subject to the per-readout density requirement (Jang et al., 21 Mar 2025).

5. Integration with Reconstruction Algorithms and Empirical Performance

Recent work integrates variable-density spiral sampling with deep learning-based reconstruction methods to achieve ultrafast and high-fidelity imaging. For instance, (Chan et al., 2024) demonstrates that optimal variable-density spirals coupled with a diffusion model-based inverse NUFFT yield SSIM > 0.87 for 23 interleaves and 20 ms acquisition, whereas uniform spirals (α=1) reach only ~0.72 and show prominent global aliasing. The denser low-frequency coverage enabled by variable density significantly reduces the conditioning requirements on the reconstruction model, allowing improved preservation of image contrast and reduction of coherent artifacts. This holds especially under high acceleration (strong undersampling) scenarios.

The HyperSLICE study (Jaubert et al., 2023) further shows that jointly optimizing the spiral density parameters and the deep artifact-suppression network, using HyperBand for automated design, yields both higher overall SSIM (~0.87 versus ~0.80 for uniform spirals) and much improved temporal fidelity after abrupt scanplan changes. Table 1 in (Jaubert et al., 2023) reports

• Optimized spiral + FastDVDnet: NRMSE = 0.127±0.026, PSNR = 32.0±2.7 dB, SSIM = 0.869±0.047
• Radial + FastDVDnet: NRMSE = 0.148±0.033, SSIM = 0.828±0.061
• Uniform-spiral + FastDVDnet: NRMSE = 0.171±0.028, SSIM = 0.802±0.054

3D implementations, as in (Jang et al., 21 Mar 2025), reveal that discretized analytic coordinate formulations for stack-of-spirals and cones require 3.8% fewer readouts for full volumetric scans at the same spatial resolution and that variable-density in both in-plane and slice directions was essential to eliminate z-aliasing under extreme undersampling.

6. Experimental Comparisons and Practical Recommendations

Extensive experimental evaluation across phantom, retrospective, and prospective in vivo acquisitions consistently demonstrates the superiority of variable-density spiral trajectories. Key findings include:

• For 3D Cones, analytic-coordinate-based variable-density trajectories reduce necessary readouts (8,862 vs 9,210) at 1.2 mm isotropic (Jang et al., 21 Mar 2025).
• With only 32 readouts (iNAV), variable-density cones (α ≈ 2.25) nearly eliminate z-aliasing, while uniform distributions yield pronounced artifacts.
• In spherical stack-of-spirals coronary MRI, combined radial (α_r=1.5) and slice (α_z=2.25) density modulation visually sharpens edges and reduces streaking.
• In interactive cardiac imaging, HyperSLICE reaches SSIM 0.869±0.047 versus 0.802±0.054 (uniform spiral) and 0.828±0.061 (radial), especially benefiting transition frames after abrupt spatial changes (Jaubert et al., 2023).

Tables from (Jaubert et al., 2023):

Trajectory	NRMSE	PSNR (dB)	SSIM	LAPE
Radial + FD-FastDVDnet	0.148±0.033	30.7±2.8	0.828±0.061	0.487±0.111
Uniform spiral + FD	0.171±0.028	29.3±2.4	0.802±0.054	0.540±0.099
Optimized spiral + FD	0.127±0.026	32.0±2.7	0.869±0.047	0.591±0.101

Recommendations in (Jaubert et al., 2023) specify $\rho(r) \propto r^{1/\alpha - 1}$6 with a Hanning transition and ~15 interleaves for cardiac bSSFP at 1.5 T.

7. Limitations and Future Directions

While variable-density spiral sampling robustly provides improved image quality under high acceleration, it incurs several limitations:

• Offline optimization (e.g., HyperBand tuning) is computationally intensive, requiring up to six days for network–trajectory joint selection (Jaubert et al., 2023).
• Once optimized, the pipeline is rigid with respect to TR, FOV, and coil geometry, though moderate generalization is possible.
• The acquisition is sensitive to system imperfections, including trajectory errors from off-resonance and gradient nonlinearities. Mitigation may require real-time trajectory correction (Jaubert et al., 2023).
• Discretization artifacts may arise for small numbers of interleaves, though analytic-coordinate discretization produces smooth, well-distributed trajectories even in these regimes (Jang et al., 21 Mar 2025).

A plausible implication is continuing research integrating real-time trajectory measurement, further adaptive density modulation, and deep joint optimization of trajectory-reconstruction pairs to address these limitations and expand clinical robustness.

---

References:

• "Design of 3D Non-Cartesian Trajectories for Fast Volumetric MRI via Analytic Coordinate Discretization" (Jang et al., 21 Mar 2025)
• "Learning the Domain Specific Inverse NUFFT for Accelerated Spiral MRI using Diffusion Models" (Chan et al., 2024)
• "Variable density sampling with continuous trajectories. Application to MRI" (Chauffert et al., 2013)
• "HyperSLICE: HyperBand optimized Spiral for Low-latency Interactive Cardiac Examination" (Jaubert et al., 2023)