Abstract:

When simultaneous multi-slice (SMS) imaging is combined with in-plane undersampling, reconstruction must unfold superimposed slices and remove aliasing simultaneously from a highly folded, undersampled K-space. To address the fixed-strength data consistency and the prior-scan mismatch in existing diffusion-based reconstructions, an SMS reconstruction method combining geometry-gated data consistency with test-time low-rank adaptation is proposed. At each reverse sampling step, the data-consistency correction is decomposed with respect to the local tangent space of the denoiser: the tangential component is retained in full, whereas the normal component is admitted through an adaptive gate, suppressing off-manifold artefacts. Low-rank branches are further attached to the score network and updated at inference using a measurement-consistency loss alone, with the backbone frozen. Experiments on simulated fastMRI brain T₂-weighted data show that the proposed method outperforms five comparison methods under all conditions, with PSNR gains of 2.27~5.45 dB over the second-best method and a more graceful degradation as acceleration increases.

Key words: Magnetic resonance imaging, Simultaneous multi-slice imaging, Diffusion model, Deep learning, Test-time adaptation