Skip to main content

Scattering-Aware Implicit Neural Representation for CT Metal Artifact Reduction

By
Ruibo Wang; Junran Yang; Ziyi Shen; Zhangling Chen; Ce Wang; Dong Liang; Kun Shang

Implicit neural representation (INR) has emerged as a promising unsupervised solution to CT metal artifact reduction (MAR). However, residual halo artifacts often persist near metallic implants because attenuation-only forward models neglect in-scatter and thus introduce a systematic forward-model mismatch. In this paper, we propose SA-INR, a scattering-aware INR framework that augments the conventional polychromatic attenuation model with a lightweight in-scatter correction inspired by the photon transport equation. SA-INR admits both a physics-constrained version and a learnable version, enabling a trade-off between interpretability and flexibility. Experiments on DeepLesion show that SA-INR improves MAR quality and robustness. Notably, the physics-constrained version improves over the state-of-the-art INR baseline without introducing additional trainable parameters, while the learnable version achieves the best overall performance (34.37 dB PSNR, 0.9499 SSIM). A capacity-matched ablation further confirms that the gain stems from the scattering-aware prior rather than from additional learnable parameters. Furthermore, we show that a strongly forward-peaked scattering prior is essential for stable optimization. These results indicate that modest but physically structured forward-model corrections can improve both reconstruction quality and robustness in unsupervised INR-based CT MAR.

Read on IEEE Xplore