Dynamic B0 and Static RF Kernels in MRI

• Dynamic B₀ and static RF kernels are spatial encoding constructs that use time-varying magnetic fields and fixed coil sensitivities to enable non-Fourier MRI encoding.
• The framework employs GRAPPA-style auto-calibration and group–patch compression to reduce computational costs and memory usage, achieving up to an 18× reduction in model size.
• Experimental results show high-quality, artifact-free reconstructions with acceleration factors up to 14.6× in 3D imaging, demonstrating practical benefits for rapid MRI.

Dynamic B₀ and Static RF Kernels

Dynamic B₀ and static radiofrequency (RF) kernels are fundamental spatial encoding constructs that enable highly accelerated magnetic resonance imaging (MRI) using joint manipulation of time-varying magnetic fields (B₀) and spatially variant coil receive sensitivities (RF). In emerging MRI strategies such as Wave-CAIPI, FRONSAC, and local B₀ coil modulations, dynamic B₀ kernels (time-dependent spatial phase maps) are imposed during the oversampled readout to encode additional degrees of freedom. Static RF kernels, representing fixed coil sensitivity patterns, are long established in parallel imaging. However, the joint use of these two kernel types introduces a non-Fourier-encoded dimension that is not amenable to conventional fast Fourier transform (FFT) inversion, motivating the development of new models, calibration, and compression approaches to maintain encoding efficiency while keeping computational costs tractable (Tian et al., 8 Jan 2026, Tian et al., 9 Jan 2026).

1. Joint Dynamic B₀ / Static RF Encoding Formalism

The MRI signal from coil $n$ at time $t$ is modeled as

$$S_n(t) = \int C_n(\mathbf{r})\, \rho(\mathbf{r})\,\exp\bigl[-i\,\mathbf{k}(t)\cdot\mathbf{r} - i\,\phi_m(t,\mathbf{r})\bigr]\,d\mathbf{r}$$

where $C_n(\mathbf{r})$ is the static (time-invariant) RF receive sensitivity, $\mathbf{k}(t)$ encodes the conventional Fourier trajectory, and

$$\phi_m(t, \mathbf{r}) = \gamma \sum_j B_j(\mathbf{r})\,h_j(t)$$

gives the accrued phase due to dynamic B₀ modulation—implemented via time-varying local coils or gradient elements with spatial profiles $B_j(\mathbf{r})$ and temporal waveforms $h_j(t)$. After discretization, the composite encoding matrix

$$E_{(n,t),r} = C_n(\mathbf{r})\,\exp\left[ -i\,\mathbf{k}(t)\cdot\mathbf{r} \right]\,\exp\left[ -i\,\phi_m(t,\mathbf{r}) \right]$$

acts on the image vector. In this construction, static RF kernels correspond to $C_n(\mathbf{r})$, while dynamic B₀ kernels are the phase terms $\exp\bigl[-i\,\phi_m(t,\mathbf{r})\bigr]$ that vary with $t$ (Tian et al., 8 Jan 2026, Tian et al., 9 Jan 2026).

2. Signal Model, Kernel Calibration, and Patch Grouping

Convolutional modeling in $k$-space reveals that the modulated signal can be expressed as

$$s_i(\mathbf{k}, t) = [C_i(\mathbf{r})]^\wedge \ast [e^{-i\phi(\mathbf{r}, t)}]^\wedge \ast S(\mathbf{k})$$

with $S(\mathbf{k})$ the unmodulated image spectrum. To estimate the dynamic B₀ kernels, generalized GRAPPA-style auto-calibration is employed: two auto-calibration signal (ACS) regions are acquired—one standard and one with added B₀ modulation. The time-indexed kernel $W^{(t)}$ solves

$$s_{\text{mod}}(k, t) = \sum_{p, n} W^{(t)}_{p, n} s_{\text{std}}(k+p, t-n)$$

forming a linear system solved by SVD, LSQR, or pseudoinverse. For experimental B₀ modulations that are periodic or repeatable, multiple $k$-space points share identical phase modulations, meaning that kernels $W_m$ can be grouped by unique phase patterns over a modulation cycle (Tian et al., 9 Jan 2026).

3. Group-Patch Compression and Joint Kernel Subspaces

To address the prohibitive computational growth of the encoding matrix, group–patch joint compression is leveraged, acting on the expanded coil × time axis. Patches of $k$-space are grouped such that each group $g$ is associated with a unique set of dynamic B₀ and RF modulation patterns. A representative composite encoding matrix $E_g$ for group $g$ combines conventional gradient ($E_g^{LG}$), dynamic B₀ ($E_g^{B0}$), and static RF ($E^{RF}$) maps as

$E_g = E_g^{LG} \odot E_g^{B0} \odot E^{RF}$

where $\odot$ denotes elementwise multiplication.

For each group, the encoding covariance $P_g = E_g N_g E_g^H$ (with pixel normalization $N_g$) is subjected to SVD, and a subspace of $N_s' \ll N_s$ modes is selected based on energy retention (typically 90–95%). The resulting compression matrix $A_g = C_g U_g^H T$ (with optional noise whitening $T$) projects both data and encoding maps onto a compact subspace, producing virtual channels that are mixtures of coils and timepoints. $A_g$ is reused for all patches in group $g$, requiring only one SVD per group and dramatically reducing memory and computational costs (Tian et al., 8 Jan 2026).

4. Patchwise k-space Reconstruction and RKHS Formulation

After kernel calibration and compression, subregion-wise or patchwise $k$-space reconstruction is performed. For each subregion $R_j$, which spans an integer number of B₀ modulation periods, source encoding matrices $E_{\rm src}$ and target encoding matrices $E_{\rm tgt}$ are formed: $$E_{\rm src} = [c_\ell(\mathbf{r}) e^{-i\phi_m(\mathbf{r})} e^{-i2\pi \mathbf{k} \cdot \mathbf{r}}]_{\,\mathbf{r},\,(\,\ell,t\,)}, \quad E_{\rm tgt} = [e^{-i2\pi \mathbf{k}' \cdot \mathbf{r}}]_{\,\mathbf{r},\,k'\in R_j}$$ The cardinal mapping matrix $U$ is found by solving $M U = R$ with $M = E_{\rm src}^H E_{\rm src}$ and $R = E_{\rm src}^H E_{\rm tgt}$. Applying $U^H$ to the acquired source data yields the fully-sampled target data for that subregion. The structure of the method ensures that kernels sharing instantaneous image-space modulation are used repeatedly, and robust reconstruction is maintained even in the presence of eddy currents and hardware imperfections (Tian et al., 9 Jan 2026).

5. Quantitative Performance, Computational Gains, and Limits

Joint compression of dynamic B₀ and static RF kernels attains compression factors of 11×–20× with negligible encoding-efficiency loss (NRMSE ≲ 0.6%, SSIM > 0.95). For a 3D 180×180×80 dataset with 32 coils and 8× oversampling, RF-only compression achieves a 3.2× reduction (matrix size: 93 GB, CS-recon: 717 s) versus a joint compression of 18× (matrix size: 53 GB, CS-recon: 177 s). In comparative studies, artifact-free images were recovered under both linear and nonlinear B₀ modulations, including Wave-CAIPI and FRONSAC, with acceleration factors up to 8× (2D) and 14.6× (3D), and per-slice/volume reconstructions in the 1.4–5.1 s (2D) and 177 s–10.1 min (3D) range, respectively. Grouped calibration and reconstruction deliver substantial reductions in both computational demand and memory footprint—e.g., 300 GB to 50 GB forward-model size for 180×180×80 scans (Tian et al., 8 Jan 2026, Tian et al., 9 Jan 2026).

6. Assumptions, Extensions, and Limitations

Joint dynamic B₀ / static RF kernel approaches presume that B₀ modulations are periodic or sufficiently repeatable for group-wise calibration. Accurate auto-calibration is essential; severe mismodeling of phase maps will degrade reconstructions. Grouping requires that the modulation patterns can be partitioned into a finite set of states or phase bins, typically achievable for sinusoidal or programmed local coil waveforms.

Extensions are possible to multi-frequency or hybrid encodings (e.g., readout-CAIPI, dual-polarity EPI), non-Cartesian trajectories, and the integration of additional encoding functions (e.g., subpixel encoding, RF transmission variation). Alternative compression mechanisms—including learned neural compressors and RKHS-based interpolations—are compatible within the group–patch framework. Combining these joint strategies with GPU acceleration or learned patchwise reconstruction may provide further gains. The primary limiting factors remain calibration fidelity, the need for groupwise modulation stationarity, and the invertibility of patchwise systems as group/patch sizes grow (Tian et al., 8 Jan 2026, Tian et al., 9 Jan 2026).