[Paper Review] Back to Physics: Operator-Guided Generative Paths for SMS MRI Reconstruction

Chen, Zhibo, Yu Guan+6 more|arXiv (Cornell University)|Feb 8, 2026

Advanced Neuroimaging Techniques and Applications0 citations

TL;DR

The paper introduces an operator-guided, physics-informed reconstruction framework for SMS MRI that uses an operator-conditional dual-stream network (OCDI-Net) to model deterministic degradation from SMS coil-slice interference and in-plane undersampling, enabling two-stage inference for improved fidelity and reduced slice leakage.

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.

Motivation & Objective

• Motivate a reconstruction paradigm that treats SMS degradation as deterministic, operator-driven processes rather than stochastic noise.
• Develop an operator-guided generative framework that aligns the inversion trajectory with acquisition operators.
• Propose OCDI-Net to disentangle target slice content from inter-slice interference and predict operator-induced degradations.
• Decouple SMS slice separation from in-plane data completion via a two-stage inference pipeline.
• Demonstrate improved fidelity and reduced slice leakage on fastMRI brain data and prospectively acquired diffusion MRI data.

Proposed method

• Formulate SMS reconstruction as deterministic degradation governed by known acquisition operators CAIPI slice modulation and Cartesian undersampling and construct a stage-specific degradation path.
• Introduce OCDI-Net, a dual-stream (target-content and interference) U-Net that predicts operator-induced degradations conditioned on diffusion step t and stage  M or U, enabling stage-aware inverse updates.
• Define a two-stage inference: Stage-M for slice separation and Stage-U for in-plane completion, both using operator-guided reverse updates with learned _t predictions.
• Use a forward deterministic trajectory x_t = k* + α_t d_Ω and reverse updates x_{t-1} = x_t - α_t d̂_t, where d̂_t is predicted by OCDI-Net.
• Train with an L1 loss on k-space reconstructions and optional image-domain loss to encourage faithful coil-combined magnitudes.
• Leverage a two-stage inference to stabilize optimization under high acceleration and allow cross-stage data consistency via pseudo-measurements and low-frequency anchors.

Experimental results

Research questions

• RQ1 Can operator-guided, physics-informed diffusion-like inference improve SMS MRI reconstructions compared to stochastic diffusion-based methods and conventional SMS reconstructions?
• RQ2 Does explicitly modeling CAIPI-induced inter-slice interference and in-plane undersampling as deterministic degradations improve slice leakage suppression and high-frequency preservation?
• RQ3 Can a two-stage inference (slice separation followed by in-plane completion) yield more stable reconstructions under high MB factors and acceleration?
• RQ4 Does OCDI-Net effectively disentangle target content from interference to enhance fidelity on retrospective fastMRI brain data and prospective TJU DWI data?

Key findings

• The proposed operator-guided framework yields improved reconstruction fidelity and reduced slice leakage compared with conventional and learning-based SMS baselines on fastMRI brain data and prospective DWI data.
• OCDI-Net effectively disentangles target slice content from structured inter-slice interference and predicts operator-induced degradations for physics-aligned inversion.
• A two-stage inference decouples slice separation and in-plane completion, enhancing stability under high acceleration and improving consistency across MB factors and R values.
• Experiments include retrospective-fastMRI simulations with MB=3 and various in-plane accelerations, plus prospective DWI data (MB=2, R=2, b=1000) showing favorable performance against baselines such as SENSE, Slice-GRAPPA, RAKI, and diffusion-based SMS methods.