Unsupervised CT Metal Artifact Reduction Via Consistent Artifact Modeling

Metal artifacts severely corrupt CT images, masking anatomical structures and undermining diagnostic reliability. Although supervised metal artifact reduction (MAR) achieves strong suppression, its dependence on paired clean and corrupted data confines it to simulated domains and hinders clinical translation. We recast unsupervised MAR as an artifact-consistent residual learning problem and propose a physically grounded, attention-guided generative framework that closes the loop between artifact estimation and removal. A single-generator dual-branch architecture estimates metal-induced corruption as a transferable additive residual and reinjects it to synthesize controlled degraded samples during cyclic training, explicitly constraining the removable component. Guided by interpolated pseudo-priors, a difference-aware attention mechanism concentrates correction on corrupted regions while safeguarding intact anatomy, all without paired supervision. Extensive experiments on synthetic benchmarks and real clinical data show that our method restores structural fidelity, suppresses streaks and beam-hardening, and generalizes robustly across diverse implants and anatomical sites. Beyond perceptual quality, it delivers state-of-the-art pelvic bone segmentation under metal corruption at efficient inference cost. This work charts an interpretable, efficient, and clinically scalable route toward reliable CT metal artifact suppression in real-world practice. Our source code is available at https://github.com/wonderfulpoet/CAM-Net.