Guidance-Coupled Nullspace Drifting for One-Step Medical Image Reconstruction

Abstract

Generative reconstruction can recover high-quality images from undersampled acquisitions, but state-of-the-art generative priors commonly rely on iterative sampling. Reducing the number of iterations lowers inference cost but also removes opportunities for progressive distributional refinement, making it difficult to retain image realism without sacrificing reconstruction fidelity. Drift modeling offers an alternative by shifting progressive refinement to training, thereby enabling one-step sampling without an iterative inference trajectory. Existing drifting methods, however, do not account for the imaging operator when modeling distributional drift. We propose DriftRecon, a generative method for one-step medical image reconstruction that performs drifting directly in the nullspace of the imaging operator. DriftRecon uses nullspace-restricted noise to construct training samples and a multiscale generator that incorporates measurement conditioning across resolution levels, with its generated component projected onto the operator nullspace. To control the fidelity--realism trade-off at inference, a conditional guidance parameter continuously steers the same generator between measurement-conditioned and unconditional generation. We derive a guidance-dependent similarity-kernel temperature schedule that preserves the guidance-weighted mass receiving strengthened supervision, thereby progressively localizing drift interactions as guidance increases to limit the associated loss of reconstruction fidelity. Experiments in sparse-view CT and accelerated MRI reconstruction show that DriftRecon outperforms established generative reconstruction methods and recent drifting models while producing reconstructions in a single forward pass.