Adaptive thresholding for CT metal artifact reduction via LangGraph

OBJECTIVE: Metal artifact reduction (MAR) in computed tomography (CT) remains a significant challenge due to inconsistencies in X-ray attenuation, which induce severe streak artifacts and obscure non-metal anatomical structures. A prevalent approach involves segmenting background and metal components within the sinogram domain. However, because the severity of artifacts varies according to the material, size, and location of the metal objects, an optimized thresholding strategy is essential to effectively distinguish background from metal regions. APPROACH: In this paper, we introduce LangGraph-MAR, a novel framework that integrates adaptive threshold optimization with automated image quality assessment. Our method utilizes a node-based LangGraph architecture, where specialized functions-including reconstruction, sinogram inpainting, and L1-based metal soft-thresholding-are assigned to discrete nodes interconnected via edges. A fundamental component of this system is the ground checking node, a quality assessment classifier trained on both artifact-corrupted and artifact-free images. This classifier facilitates an automated, iterative MAR process that converges to the optimal threshold value to ensure high image quality. MAIN RESULTS: Comparative analysis with baseline models demonstrates that LangGraph-MAR outperforms several MAR methods in CT MAR tasks, yielding SSIM relative improvements of 8.64% and 5.61% for body and head protocols against InDuDoNet+, respectively. Additional representative cases and region of interest level analyses further support the efficacy of the proposed framework, showing substantial recovery of structural information in regions affected by artifacts. SIGNIFICANCE: The proposed framework ensures consistent performance across diverse metal implants through adaptive thresholding. Furthermore, its modular plug-and-play design improves diagnostic reliability, facilitating more effective clinical decision-making. Code: Available upon acceptance.