Improving the efficiency of normalized metal artifact reduction via a unified forward projection.

Authors

Abstract

Normalized metal artifact reduction (NMAR) is a robust and widely used method for reducing metal artifacts in computed tomography (CT). However, conventional NMAR requires at least two forward projections, one for metal trace detection and the other for prior sinogram generation, resulting in redundant computation and limited efficiency. This study aims to reformulate NMAR into a single forward projection-based framework that maintains artifact reduction performance while improving computational efficiency and structural simplicity.
Approach: We show that the two separate forward projections in NMAR can be unified into a single operation by leveraging deep learning (DL) priors, thereby eliminating the explicit forward projection for metal trace. The metal trace is inferred directly from localized discrepancies between the original sinogram and the forward projection of the DL prior image, allowing both interpolation and trace identification within a unified forward projection. Simulations and cadaver experiments were performed to compare the proposed method with NMAR, DL reconstruction, and conventional DL-NMAR.
Main results: The proposed method reduced metal artifacts with image quality comparable to conventional DL-NMAR while improving computational efficiency. By reducing the number of forward projections from two to one, the proposed method achieved the lowest number of projection operations among all compared methods, highlighting its computational advantage.
Significance: This study demonstrates that deep learning priors can be seamlessly integrated into physics-based NMAR frameworks to simplify image reconstruction process and enhance computational performance. The proposed unified forward projection provides an efficient solution to accelerate metal artifact reduction in CT imaging.

Tags

Topics