Back to Physics: Operator-Guided Generative Paths for SMS MRI Reconstruction

TL;DR

This work introduces an operator-conditional dual-stream interaction network (OCDI-Net) that explicitly disentangles target-slice content from inter-slice interference and predicts structured degradations for operator-aligned inversion, and instantiate reconstruction as a two-stage chained inference procedure that performs SMS slice separation followed by in-plane completion.

Abstract

Simultaneous multi-slice (SMS) imaging with in-plane undersampling enables highly accelerated MRI but yields a strongly coupled inverse problem with deterministic inter-slice interference and missing k-space data. Most diffusion-based reconstructions are formulated around Gaussian-noise corruption and rely on additional consistency steps to incorporate SMS physics, which can be mismatched to the operator-governed degradations in SMS acquisition. We propose an operator-guided framework that models the degradation trajectory using known acquisition operators and inverts this process via deterministic updates. Within this framework, we introduce an operator-conditional dual-stream interaction network (OCDI-Net) that explicitly disentangles target-slice content from inter-slice interference and predicts structured degradations for operator-aligned inversion, and we instantiate reconstruction as a two-stage chained inference procedure that performs SMS slice separation followed by in-plane completion. Experiments on fastMRI brain data and prospectively acquired in vivo diffusion MRI data demonstrate improved fidelity and reduced slice leakage over conventional and learning-based SMS reconstructions.

Figures (6)

• Figure 1: Conceptual comparison of SMS reconstruction paradigms. (a) Linear operator-based methods recover slices from collapsed measurements. (b) Diffusion-based approaches combine unaliasing with stochastic denoising and iterative refinement. (c) Our operator-guided method predicts operator-induced components for deterministic inversion. KMB is the collapsed measurement, K1 is the target slice, and K2+K3 denotes the predicted interference removed to recover K1.
• Figure 2: Architecture of the operator-conditional dual-stream interaction network (OCDI-Net). OCDI-Net maintains a target-content stream and an interference stream throughout a U-Net-style encoder--decoder r15. Cross-stream interaction blocks exchange information under conditioning on diffusion step $t$ and operator/stage indicator $\Omega$. The network outputs the predicted degradation $\widehat{\mathbf{d}}_t$ used by the operator-guided reverse updates in Sec. \ref{['sec:method_cold_formulation']}.
• Figure 3: Overall pipeline of the proposed operator-guided SMS reconstruction framework. $K_1$, $K_2$, and $K_3$ denote individual slice k-space components, and $K_{\mathrm{MB}}$ denotes their CAIPI-modulated multiband superposition. The suffix "us" indicates in-plane undersampled k-space. During reconstruction, each target slice is aligned by inverse CAIPI modulation and reconstructed by two chained stages using the same OCDI-Net: Stage-M performs slice separation, followed by Stage-U for in-plane completion.
• Figure 4: Visual comparisons on fastMRI under MB3R2. Columns show GT and reconstructions by SENSE, Slice-GRAPPA (SG), RAKI, ROGER, and the proposed method. Each example shows the full-FOV RSS image with ROI and the zoomed ROI with absolute error map.
• Figure 5: Quantitative comparison on fastMRI under MB3 with varying in-plane acceleration factors. Left: PSNR. Right: SSIM.NMSE trends are consistent and omitted for clarity.