SCMA flow matching removes CT metal artifacts while preserving anatomy

Metal implants in X-ray CT cause beam hardening, photon starvation, and scattering, producing streaks and dark bands that can obscure diagnosis. Previous metal artifact reduction (MAR) methods struggled: optimization approaches left residual blur, regressions generalized poorly, and generative models often generated anatomically inconsistent structures. SCMA, proposed by Wang et al., leverages flow matching—a generative technique that learns a velocity field to deterministically map noise to a clean distribution—but improves on vanilla flow matching with three key innovations.

First, SCMA feeds a linear-interpolation-corrected image into the velocity network alongside the intermediate state, giving sample-specific structural guidance that preserves anatomy. Second, it incorporates time-varying spatial weights derived from the metal mask and distance transform into the flow-matching loss, forcing the model to focus on severe degradation near metal regions. Third, it alternates conditional flow-matching updates with projection-consistency correction, so measurements outside metal traces constrain predictions. Evaluations on both simulated and real CT data show SCMA suppresses metal artifacts better than representative MAR methods while preserving local structures and reducing hallucination-like artifacts inconsistent with measured projections.