k-Space Correction in MRI

• k-space correction is a method that modifies and regularizes the Fourier domain data to ensure MRI reconstructions remain faithful to the underlying acquisition physics.
• Techniques include diagonal preconditioning, self-supervised regularization (e.g., PISCO), and learned residual corrections like SPARK, each targeting specific artifacts from nonuniform sampling or motion errors.
• These methods adjust optimization dynamics or the reconstruction objective, providing enhanced convergence and image quality while compensating for hardware imperfections and model mismatches.

Searching arXiv for recent and foundational papers on k-space correction in MRI and related computational models. k-space correction denotes a class of methods that modify, regularize, or compensate data in the spatial-frequency domain of magnetic resonance imaging and related wave-based imaging models so that subsequent reconstruction is more faithful to the underlying acquisition physics. In MRI, these methods address disparate failure modes that originate directly in k-space, including nonuniform sampling ill-conditioning, motion-induced inconsistency between lines or spokes, trajectory deviations, phase mismatches, field inhomogeneity, partial Fourier truncation, and interpolation errors in accelerated or dynamic acquisitions. Across these settings, the common principle is that the correction is imposed before or during inversion of the forward model, either by reconditioning the optimization, enforcing self-consistency relations, estimating residual error terms, or explicitly modeling phase and sampling perturbations in the encoding operator. Representative formulations include diagonal k-space preconditioning for non-Cartesian iterative reconstruction (Ong et al., 2019), self-supervised k-space regularization via parallel imaging-inspired self-consistency (PISCO) for neural implicit k-space representations (Spieker et al., 16 Jan 2025), reference-free k-space deep learning for EPI ghost correction (Lee et al., 2018), structured low-rank recovery for EPI $B_0$ correction (Balachandrasekaran et al., 2018), and scan-specific residual correction in parallel MRI (Beker et al., 2019).

1. Conceptual scope and problem classes

Within MRI, k-space correction is not a single algorithmic family but a unifying viewpoint: reconstruction errors are treated as k-space-domain inconsistencies whose structure can be exploited. The target inconsistency may be algorithmic, as in poor conditioning caused by variable-density non-Cartesian sampling; physical, as in gradient delay, eddy-current, off-resonance, or even/odd echo phase mismatch; or statistical, as in overfitting of learned k-space representations under aggressive undersampling. The correction may therefore enter as a preconditioner, a learned residual, a self-consistency constraint, a structured low-rank prior, or an explicit phase/distortion term in the forward model.

A key distinction is between methods that alter the reconstruction objective and methods that alter only the optimization dynamics. In non-Cartesian MRI, density compensation functions reweight the data-consistency term, solving

$$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$

which can accelerate convergence but can also bias reconstructions and color noise. By contrast, k-space preconditioning in the dual formulation changes the update rule while preserving the exact unweighted objective (Ong et al., 2019). This distinction recurs elsewhere: SPARK predicts a residual error term on reconstructed coil k-spaces rather than replacing the physics-based reconstructor (Beker et al., 2019), while PISCO regularizes a neural implicit k-space model without requiring external calibration data (Spieker et al., 16 Jan 2025).

Another important distinction is between hardware-related correction and representational regularization. Hardware-related methods correct encoding imperfections such as phase mismatch in EPI (Lee et al., 2018), field inhomogeneity in EPI (Balachandrasekaran et al., 2018), or trajectory errors in radial imaging (Mani et al., 2018). Representational methods instead constrain what a plausible k-space should look like under undersampling, parallel imaging redundancy, or learned implicit parameterization (Spieker et al., 2024, Spieker et al., 16 Jan 2025). This suggests that “k-space correction” encompasses both compensation of acquisition errors and enforcement of reconstruction-consistent frequency-domain structure.

2. Optimization-based correction: preconditioning and conditioning control

A canonical optimization-based formulation appears in non-Cartesian MRI reconstruction with forward model

$y = A x + w,$

where $A = M F_u S$, $S$ stacks coil sensitivities, $F_u$ is a nonuniform Fourier operator, and $M$ is a sampling operator (Ong et al., 2019). For objectives of the form

$\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$

the ill-conditioning of $A^H A$ under variable-density sampling slows first-order and splitting methods. The contribution of "Accelerating Non-Cartesian MRI Reconstruction Convergence using k-space Preconditioning" is to move to the dual problem, where the dual variable $u$ lives in data space, and to precondition directly in k-space using a diagonal matrix $$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$0 (Ong et al., 2019).

With primal-dual hybrid gradient, the preconditioned updates become

$$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$1

$$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$2

with analogous extensions for composite penalties such as total variation. The element-wise dual update,

$$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$3

shows that each k-space sample receives location-specific step control. The diagonal weights are chosen to approximate $$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$4 in Frobenius norm, via

$$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$5

leading to closed-form expressions that reduce, in the single-coil NUFFT case, to a discrete-sinc-weighted local density measure (Ong et al., 2019).

The significance of this method is twofold. First, it preserves the exact objective rather than solving a weighted surrogate. Second, it adds only $$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$6 element-wise multiplications per iteration, unlike circulant preconditioners that require additional FFTs and typically induce inner-loop overhead (Ong et al., 2019). Empirically, the method converged in about ten iterations on challenging 3D UTE data, whereas without preconditioning extreme blurring persisted even after 100 iterations (Ong et al., 2019). For this reason, k-space correction here is best understood as conditioning correction rather than data modification.

A related but conceptually broader development is the Fourier-domain basis expansion proposed for non-Cartesian Fourier imaging (Chan et al., 8 May 2025). That work argues that standard voxel-based image-domain models suffer from approximation, periodicity, and nullspace pathologies, and replaces them with a k-space basis

$$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$7

This is not presented as a preconditioner, but it functions as a structural k-space correction by changing the discretization so that basis functions are localized and non-periodic in k-space, mitigating wrap-around leakage and improving conditioning (Chan et al., 8 May 2025). A plausible implication is that conditioning correction can be achieved not only through optimization dynamics, but also through the choice of representational basis.

3. Self-consistency and regularization in learned k-space models

In accelerated and dynamic MRI, k-space correction increasingly takes the form of self-supervised regularization. Neural implicit k-space representations model multi-coil dynamic k-space as a coordinate-based map

$$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$8

trained only on acquired coordinates. Under strong acceleration, especially in radial dynamic MRI, these models overfit the sparsely constrained periphery of k-space (Spieker et al., 16 Jan 2025). PISCO addresses this by importing the parallel imaging notion of global k-space neighborhood consistency into a self-supervised loss.

Given patch-target matrices $$\min_x \frac{1}{2}\| D^{1/2}(A x - y) \|_2^2 + g(x),$$9 and $y = A x + w,$0 from multiple random k-space subsets, PISCO solves

$y = A x + w,$1

and enforces that the resulting local linear relationships are globally consistent. The residual-based formulation defines

$y = A x + w,$2

which was found more stable than a previously used distance-based penalty on pairwise weight differences (Spieker et al., 16 Jan 2025). In practice, Cartesian $y = A x + w,$3 kernels with spacing

$y = A x + w,$4

were used, and patches within each subset were restricted to a single time point to avoid temporal blurring (Spieker et al., 16 Jan 2025).

The full training objective becomes

$y = A x + w,$5

with a preconditioning phase in which $y = A x + w,$6 for the first $y = A x + w,$7 epochs in the 2025 formulation (Spieker et al., 16 Jan 2025). Quantitatively and qualitatively, integrating PISCO improved static and dynamic reconstructions, especially at high acceleration factors $y = A x + w,$8, and yielded superior spatio-temporal reconstruction quality compared to state-of-the-art methods (Spieker et al., 16 Jan 2025). The earlier abdominal study reported gains up to $y = A x + w,$9 dB PSNR and FSIM/FSIM-t improvements up to $A = M F_u S$0 and $A = M F_u S$1, respectively (Spieker et al., 2024).

PISCO is representative of a broader shift: k-space correction need not estimate an explicit error field. It can instead regularize the learned representation toward parallel-imaging-consistent local linearity. This differs from classical GRAPPA and SPIRiT in being calibration-free and integrated into an implicit-function optimizer (Spieker et al., 16 Jan 2025, Spieker et al., 2024). It also differs from image-domain priors in that it remains entirely within the k-space training regime.

A related learned-residual paradigm is SPARK, which trains scan-specific CNNs to predict coil-wise k-space correction terms from reconstructed coil k-spaces. For each coil $A = M F_u S$2, the ACS mismatch is

$A = M F_u S$3

and the CNN is trained by

$A = M F_u S$4

The full corrected k-space is then $A = M F_u S$5 (Beker et al., 2019). SPARK reduced RMSE by up to $A = M F_u S$6-fold versus GRAPPA and up to $A = M F_u S$7-fold versus RAKI on the reported datasets (Beker et al., 2019). In encyclopedic terms, SPARK exemplifies k-space correction as residual error generalization from ACS discrepancies.

4. Correction of phase mismatch, off-resonance, and trajectory errors

A large branch of k-space correction addresses encoding imperfections that are most naturally expressed as phase inconsistency or trajectory deviation.

In single-shot EPI, Nyquist ghosting arises from phase mismatch between even and odd echoes. The alternating phase error in phase-encode line index $A = M F_u S$8,

$A = M F_u S$9

contains a dominant periodicity-2 component that produces a ghost shifted by $S$0 in the phase-encode direction (Lee et al., 2018). "k-Space Deep Learning for Reference-free EPI Ghost Correction" reformulates the problem as k-space interpolation of two virtual polarity-dependent k-spaces, leveraging structured low-rank Hankel ideas and their deep convolutional reinterpretation (Lee et al., 2018). Inputs are even/odd channels per coil, complex-split into real and imaginary parts, and the network estimates the missing virtual k-space data directly. Reconstruction results on 3T and 7T data showed lower ghost-to-signal ratio and far faster runtime than low-rank completion, with per-slice/per-frame time of approximately $S$1 ms versus $S$2 s for ALOHA (Lee et al., 2018).

Off-resonance correction in EPI can also be cast as k-space-domain structured recovery. In dual-echo EPI, each voxel time course across segmented readout times is modeled as a single exponential,

$S$3

where $S$4 encodes both $S$5 decay and $S$6 (Balachandrasekaran et al., 2018). This yields a 3D annihilation relation in $S$7, hence a low-rank multi-fold Toeplitz matrix. The resulting calibration-free method recovers the full 3D k-space volume and returns an undistorted image as the virtual frame at $S$8, without requiring an external field map (Balachandrasekaran et al., 2018). The method exploits the fact that EPI is fully sampled along $S$9, enabling null-space estimation from fully sampled rows.

Trajectory-error correction in radial imaging admits a similar low-rank treatment. If segment $F_u$0 experiences a k-space shift $F_u$1, then

$F_u$2

and the k-spaces of different segments satisfy annihilation relations that render a multi-block Hankel matrix low-rank (Mani et al., 2018). The T-MUSSELS formulation solves

$F_u$3

thereby recovering artifact-free images from radial data without trajectory calibration (Mani et al., 2018). This generalizes from constant-shift phase errors to spatially smooth inter-segment phase differences.

In simultaneous multi-slab diffusion MRI with blipped-CAIPI, correction requires simultaneous treatment of multiple k-space phase terms. In the hybrid-space REACH/DC-REACH framework, blipped-CAIPI-induced ramp phase $F_u$4, motion phase $F_u$5, background phase $F_u$6, and distortion phase $F_u$7 are embedded in a single forward operator,

$F_u$8

so that reconstruction jointly corrects z–km phase interference, motion, and EPI distortion (Zhang et al., 2023). The reported joint reconstruction improved mean $F_u$9-factor by around $M$0 in blipped-SMSlab experiments (Zhang et al., 2023).

These examples show that when the physical error mechanism is known, k-space correction is most effective when the corresponding phase or trajectory perturbation is absorbed into the encoding model rather than post hoc suppressed in the image domain.

5. Partial Fourier, ghost, and motion correction as k-space-domain completion

Partial Fourier and motion corruption provide two additional settings in which k-space correction is framed as completion or replacement of inconsistent samples.

For partial k-space acquisition, higher-dimensional homodyne filtering extends classical 1D homodyne phase correction to truncation along multiple encoding directions (Paul et al., 2015). In 2D, separate ramp weightings $M$1 and $M$2 are applied, low-resolution phase is estimated per axis, and the corrected image is formed by averaging the real parts of the separately demodulated reconstructions. In 3D, the weighted reconstructions from the three truncated axes are summed before demodulation with a phase estimated from the summed low-resolution masks (Paul et al., 2015). This procedure was reported to suppress incidental phase artifacts while introducing less signal loss than conventional homodyne (Paul et al., 2015). The method is best regarded as k-space correction through asymmetric-sampling bias compensation and low-resolution phase demodulation.

An older but related line of work uses explicit FIR prediction models for phase-error compensation and artifact suppression in partial k-space. In phase-encode direction, frequency-weighted signal-space prediction with subspace projection extrapolates missing negative lines while compensating phase errors; in frequency-encode direction, least-squares FIR models in an intermediate Fourier domain extrapolate truncated echoes and suppress Gibbs ringing and streaking (Paul et al., 2013). The reported method achieved artifact-free reconstruction while using only half the number of fractional lines required by some existing methods, according to the paper’s comparison with homodyne-type baselines (Paul et al., 2013).

Motion corruption in Cartesian cine CMR has been treated by first detecting corrupted $M$3-$M$4 lines and then reconstructing from the remaining lines as an undersampled problem. The corrupted measurement model is

$M$5

with $M$6 indicating corrupted lines (Oksuz et al., 2019). A detection network estimates the corruption mask, after which a recurrent reconstruction network enforces data consistency only on uncorrupted lines. On a test set of 50 cine datasets, the end-to-end method achieved PSNR $M$7, RMSE $M$8, and SSIM $M$9, while preserving clean images with PSNR $\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$0, RMSE $\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$1, and SSIM $\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$2 (Oksuz et al., 2019). Here k-space correction means identifying and nullifying inconsistent lines so that the inverse problem is reformulated with an acquisition-consistent mask.

SISMIK pursues a different motion strategy: it estimates per-line rigid-body motion directly from complex k-space windows using CNNs, then reconstructs with a motion-aware forward model and NUFFT (Dabrowski et al., 2023). The translational part is corrected by inverse line-wise phase ramps, while rotations are handled by nonuniform regridding. On simulations, SISMIK plus NUFFT achieved median PSNR $\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$3 dB and SSIM $\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$4, compared with $\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$5 dB and $\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$6 for GradMC (Dabrowski et al., 2023). This again illustrates the general pattern that k-space correction often consists of identifying inconsistent subsets and re-estimating the image under an updated encoding model.

6. Broader methodological landscape and non-MRI usage

Although MRI dominates the modern use of the phrase, k-space correction also appears in wave-propagation simulation. In k-space pseudospectral time-domain models of acoustics, time stepping induces numerical dispersion even when spatial derivatives are exact in the Fourier domain. For uniform time step $\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$7, this is removed by the correction factor

$\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$8

which multiplies the spectral gradient terms and yields numerically exact propagation in a homogeneous medium (King et al., 10 Jul 2025). For non-uniform time stepping, the 2025 work derives a two-term correction $\min_x \frac{1}{2}\|A x - y\|_2^2 + R(x),$9 that bridges unequal half-step stencils and avoids spurious reflections (King et al., 10 Jul 2025). This usage is not about MRI reconstruction, yet it preserves the central meaning of k-space correction: compensating Fourier-domain discretization error without altering the underlying physical solution.

Across MRI and acoustics, three recurrent abstractions emerge.

Abstraction	Mechanism	Representative papers
Conditioning correction	Diagonal or structural modification of updates/basis to improve inversion stability	(Ong et al., 2019, Chan et al., 8 May 2025)
Consistency enforcement	Low-rank, self-consistency, or residual constraints in k-space	(Spieker et al., 16 Jan 2025, Beker et al., 2019, Balachandrasekaran et al., 2018, Mani et al., 2018)
Physics compensation	Explicit modeling of phase, trajectory, or time-stepping perturbations	(Lee et al., 2018, Zhang et al., 2023, Dabrowski et al., 2023, King et al., 10 Jul 2025)

This suggests that “k-space correction” is less a specific algorithmic primitive than a domain choice: errors are corrected where they arise, namely in the sampled Fourier representation or in its immediate optimization dual.

7. Limitations, misconceptions, and research directions

A common misconception is that k-space correction is synonymous with density compensation or regridding. The literature shows otherwise. Density compensation changes the objective and may sacrifice accuracy (Ong et al., 2019); regridding only handles nonuniform sample placement and does not by itself correct phase mismatch, motion corruption, or learned overfitting. Conversely, some methods described as “correction” are regularizers rather than explicit error estimators, as in PISCO (Spieker et al., 16 Jan 2025). The term therefore spans both compensatory and prior-enforcing operations.

Another misconception is that k-space-domain methods necessarily require calibration data. Several influential methods are calibration-free or scan-specific without external references: PISCO does not require ACS (Spieker et al., 16 Jan 2025); structured low-rank EPI $A^H A$0 correction estimates its null space from the acquired data (Balachandrasekaran et al., 2018); T-MUSSELS corrects trajectory errors without trajectory calibration (Mani et al., 2018); SPARK learns scan-specific residuals from ACS mismatch without an external training set (Beker et al., 2019). This suggests that calibration-free k-space correction is a mature subtheme rather than an exception.

The main technical limitations are equally recurrent. Many methods assume smooth phase or smooth parameter fields, as in structured low-rank off-resonance correction (Balachandrasekaran et al., 2018) and segment-based trajectory correction (Mani et al., 2018). Learned methods depend on kernel geometry, training distribution, or coil count scaling, as seen in PISCO’s sensitivity to Cartesian kernel choice and $A^H A$1 cost (Spieker et al., 16 Jan 2025). Preconditioning methods depend on accurate forward models and do not fix model mismatch such as trajectory miscalibration; those must already be encoded in $A^H A$2 (Ong et al., 2019). Motion-aware approaches may fail under through-plane or nonrigid motion not represented in the model (Dabrowski et al., 2023).

Current research directions point toward hybridization. Several papers explicitly combine k-space correction with model-based reconstruction rather than replacing it: SPARK is post-reconstruction residual correction (Beker et al., 2019), SISMIK couples k-space motion estimation to motion-aware inversion (Dabrowski et al., 2023), and REACH/DC-REACH absorb multiple correction terms into a single hybrid-space operator (Zhang et al., 2023). A plausible implication is that future systems will increasingly integrate k-space correction, learned priors, and explicit encoding-operator perturbation into unified inverse problems, rather than treating correction as a preprocessing stage.

In that broader sense, k-space correction has become a central organizing principle for modern reconstruction research: artifacts and convergence failures are addressed not after image formation, but at the level of the measured Fourier data, where sampling geometry, coil redundancy, phase evolution, and physical encoding errors remain explicit.