OpenMASC: An Open-Source Pipeline for Cross-Trajectory Metal-Aware Sampling and Correction in Accelerated MRI

Zhengyi Lu; Ming Lu; Chongyu Qu; Junchao Zhu; Junlin Guo; Marilyn Lionts; Yanfan Zhu; Yuechen Yang

cs.CVSep 26, 2026

OpenMASC: An Open-Source Pipeline for Cross-Trajectory Metal-Aware Sampling and Correction in Accelerated MRI

Authors: Zhengyi Lu, Ming Lu, Chongyu Qu, Junchao Zhu, Junlin Guo, Marilyn Lionts, Yanfan Zhu, Yuechen Yang, +6 more

Organizations: Vanderbilt University, Nashville, TN, USA · Vanderbilt University Medical Center, Nashville, TN, USA · Oak Ridge National Laboratory, Oak Ridge, TN, USA

Abstract

Metal implants corrupt MRI measurements throughout $k$ -space, yet existing accelerated MRI methods assume clean data and most metal artifact reduction approaches assume fully sampled acquisitions. No public dataset provides paired $k$ -space and images with and without metal for the same anatomy, and no framework jointly addresses artifact-aware acquisition and reconstruction across sampling trajectories. We present OpenMASC, an open-source pipeline covering the full workflow from data generation to deployment. A physics-based data generation module converts public CT volumes into paired clean and metal-corrupted MRI data in both Cartesian and radial formats. MA-VarNet, an unrolled reconstruction network with a per-cascade DC Rectifier, corrects artifacts that data-consistency steps reintroduce from corrupted measurements. A reinforcement learning agent actively selects $k$ -space readouts and co-trains with the reconstruction network through a decoupled three-stage procedure. The framework is trajectory-agnostic except for the data-consistency operator, supporting both Cartesian and radial acquisition without architectural changes. Experiments on two datasets at $4\times$ and $8\times$ acceleration demonstrate consistent improvements over conventional and learned baselines on both trajectories.

Figures & tables

Figure 1: Overview of the OpenMASC pipeline. The user selects a trajectory type (Cartesian or radial) and an acceleration factor ( $4\times$ or $8\times$ ); the toolbox handles the rest. (A) The Data Generation module converts existing CT volumes into paired clean and metal-corrupted MRI data via physics-based simulation, in either Cartesian or radial format. (B) An active sampling agent and MA-VarNet (metal-aware reconstruction) are jointly trained on the generated data. (C) At deployment, the agent iteratively selects $k$ -space readouts while MA-VarNet reconstructs, repeating until the acquisition budget is exhausted to produce an artifact-reduced image.

Figure 2: Data Generation pipeline. Stage 1: CT volumes from the autoPET dataset are automatically segmented into multi-tissue phantoms. The user interactively selects the implant site and positions the virtual prosthesis, producing paired digital phantoms (with and without metal) and six voxel-wise MR property maps. Stage 2: An open-source MRI simulator generates paired Cartesian $k$ -space and reconstructed images via a TSE sequence. Stage 3: Radial $k$ -space and images are derived from the same paired data via NUFFT using MRI-NUFFT.

Method	Simulation 4 $\times$		Simulation 8 $\times$		fastMRI 4 $\times$		fastMRI 8 $\times$
Method	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$
Rand (VD) + E2E-VarNet	0.756 $\pm$ 0.015	25.83 $\pm$ 1.75	0.747 $\pm$ 0.016	25.46 $\pm$ 1.69	0.749 $\pm$ 0.029	28.72 $\pm$ 1.33	0.679 $\pm$ 0.035	27.77 $\pm$ 1.29
Equi + E2E-VarNet	0.746 $\pm$ 0.017	25.45 $\pm$ 1.72	0.716 $\pm$ 0.022	24.07 $\pm$ 1.57	0.727 $\pm$ 0.027	28.14 $\pm$ 1.28	0.651 $\pm$ 0.032	27.00 $\pm$ 1.37
Rand (VD) + FI-VarNet	0.759 $\pm$ 0.014	26.01 $\pm$ 1.71	0.752 $\pm$ 0.015	25.67 $\pm$ 1.65	0.751 $\pm$ 0.029	28.89 $\pm$ 1.27	0.683 $\pm$ 0.035	27.97 $\pm$ 1.24
Equi + FI-VarNet	0.750 $\pm$ 0.015	25.64 $\pm$ 1.68	0.721 $\pm$ 0.021	24.31 $\pm$ 1.55	0.731 $\pm$ 0.027	28.26 $\pm$ 1.27	0.656 $\pm$ 0.033	27.10 $\pm$ 1.35
SS-DDQN + FI-VarNet	0.760 $\pm$ 0.014	26.01 $\pm$ 1.70	0.755 $\pm$ 0.014	25.81 $\pm$ 1.65	0.750 $\pm$ 0.028	28.70 $\pm$ 1.23	0.685 $\pm$ 0.036	28.08 $\pm$ 1.23

Table 1: Cartesian

k

-space results at

4\times

and

8\times

acceleration. Each method combines a sampling strategy with a reconstruction network. The simulation dataset is generated from autoPET via the Data Generation module, and fastMRI + augment uses a different implant model for cross-dataset evaluation. Best results in bold.

Figure 3: Cartesian $k$ -space results shown as violin plots. Top row: SSIM; bottom row: PSNR. Left column: simulation dataset; right column: fastMRI + augment. Each method shows paired distributions for $4\times$ (blue) and $8\times$ (orange) acceleration. OpenMASC (red, rightmost) achieves the highest median and narrowest spread across all conditions.

Figure 4: Qualitative Cartesian results on the simulation dataset at $4\times$ acceleration. Rows from top to bottom: reconstructed images with region of interest highlighted by red dashed boxes, error maps, sampling masks, and action step visualizations colored by acquisition order from Step 0 to Step 42. OpenMASC produces reconstructions closest to the ground truth with the darkest error maps.

Method	Simulation 4 $\times$		Simulation 8 $\times$		fastMRI 4 $\times$		fastMRI 8 $\times$
Method	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$
Ran + E2E-VarNet	0.651 $\pm$ 0.027	21.73 $\pm$ 1.27	0.599 $\pm$ 0.032	19.86 $\pm$ 1.29	0.687 $\pm$ 0.038	24.23 $\pm$ 2.02	0.622 $\pm$ 0.042	22.95 $\pm$ 1.98
GA + E2E-VarNet	0.663 $\pm$ 0.025	22.12 $\pm$ 1.27	0.619 $\pm$ 0.030	20.91 $\pm$ 1.21	0.698 $\pm$ 0.035	25.15 $\pm$ 1.90	0.638 $\pm$ 0.039	24.02 $\pm$ 1.88
Equi + E2E-VarNet	0.667 $\pm$ 0.027	22.35 $\pm$ 1.30	0.623 $\pm$ 0.031	21.00 $\pm$ 1.31	0.681 $\pm$ 0.035	24.84 $\pm$ 1.94	0.633 $\pm$ 0.038	23.97 $\pm$ 1.91
Rand + FI-VarNet	0.663 $\pm$ 0.020	22.63 $\pm$ 1.28	0.610 $\pm$ 0.027	20.44 $\pm$ 1.30	0.711 $\pm$ 0.034	25.34 $\pm$ 2.11	0.645 $\pm$ 0.040	23.70 $\pm$ 2.04
GA + FI-VarNet	0.680 $\pm$ 0.019	23.26 $\pm$ 1.26	0.636 $\pm$ 0.024	21.71 $\pm$ 1.22	0.724 $\pm$ 0.031	26.42 $\pm$ 1.77	0.663 $\pm$ 0.037	25.09 $\pm$ 1.93

Table 2: Radial

k

-space results at

4\times

and

8\times

Figure 5: Radial $k$ -space results shown as violin plots. Top row: SSIM; bottom row: PSNR. Left column: simulation dataset; right column: fastMRI + augment. Each method shows paired distributions for $4\times$ (blue) and $8\times$ (orange) acceleration. OpenMASC (red, rightmost) achieves the highest median and narrowest spread across all conditions.

Figure 6: Qualitative radial results on the simulation dataset at $4\times$ acceleration. Rows from top to bottom: reconstructed images with region of interest (red dashed box), error maps, radial sampling masks, and action step visualizations colored by acquisition order. OpenMASC recovers sharp structural detail near the implant that baselines distort.

Method	SSIM $\uparrow$	PSNR $\uparrow$	MSE ( $\times 10^{-3}$ ) $\downarrow$	MAE ( $\times 10^{-2}$ ) $\downarrow$
w/o Active Acq. (Rand † )	0.776 $\pm$ 0.013	27.53 $\pm$ 1.53	1.879 $\pm$ 0.691	2.312 $\pm$ 0.218
w/o Active Acq. (Equi)	0.768 $\pm$ 0.013	27.09 $\pm$ 1.52	2.081 $\pm$ 0.754	2.429 $\pm$ 0.255
w/o DCR	0.710 $\pm$ 0.022	21.45 $\pm$ 1.89	7.841 $\pm$ 3.365	3.716 $\pm$ 0.416
w/o VarNet Pretrain	0.770 $\pm$ 0.013	27.11 $\pm$ 1.48	2.059 $\pm$ 0.724	2.417 $\pm$ 0.235
w/o Decoupled Training	0.776 $\pm$ 0.013	27.47 $\pm$ 1.51	1.901 $\pm$ 0.674	2.318 $\pm$ 0.219
w/o PPO-DCR Co-train	0.776 $\pm$ 0.013	27.55 $\pm$ 1.52	1.873 $\pm$ 0.687	2.303 $\pm$ 0.219

Table 3: Ablation study on Cartesian simulation dataset at

4\times

acceleration. Each row removes one component from the full model. All paired

t

-tests vs. the full model yield

p<0.001

across

n{=}720

slices. † Low-frequency biased random sampling. Best results in bold.

Explore similar work

MGMAR: Metal-Guided Metal Artifact Reduction for X-ray Computed Tomography
Mar 13, 2026Hyoung Suk Park, Kiwan JeonComputed Tomography Image Enhancement
National Institute for Mathematical Sciences, Daejeon, 34047, Republic of Korea
Self-Auditing Residual Drifting for Pathology-Preserving Accelerated Knee MRI
Jul 2, 2026Qing Lyu, Jianxu Wang, Mohammad Kawas +2Uncertainty Quantification Image Reconstruction
Department of Radiology & Biomedical Imaging, Yale School of Medicine, New Haven, CT 06510 USA · Department of Biomedical Engineering, Rensselaer Polytechnic Institute, Troy, NY 12180 USA
Next-Acceleration-Scale Prediction for Autoregressive MRI Reconstruction
May 19, 2026Yilmaz Korkmaz, Vishal M. PatelVisual Autoregressive Models Computational Imaging
Johns Hopkins University, Baltimore MD 21218, USA

Mar 13, 2026J/K move · Enter open · S save

MGMAR: Metal-Guided Metal Artifact Reduction for X-ray Computed Tomography

Hyoung Suk Park, Kiwan Jeon

National Institute for Mathematical Sciences, Daejeon, 34047, Republic of Korea

Computed Tomography Image Enhancement

Open paper AskPDF arXiv

An X-ray computed tomography (CT), metal artifact reduction (MAR) remains a major challenge because metallic implants violate standard CT forward-model assumptions, producing severe streaking and shadowing artifacts that degrade diagnostic quality. We propose MGMAR, a metal-guided MAR method that explicitly leverages metal-related information throughout the reconstruction pipeline. MGMAR first generates a high-quality prior image by training a conditioned implicit neural representation (INR) using metal-unaffected projections, and then incorporates this prior into a normalized MAR (NMAR) framework for projection completion. To improve robustness under severe metal corruption, we pretrain the encoder-conditioned INR on paired metal-corrupted and artifact-free CT images, thereby embedding data-driven prior knowledge into the INR parameter space. This prior-embedded initialization reduces sensitivity to random initialization and accelerates convergence during measurement-specific refinement. The encoder takes a metal-corrupted reconstruction together with a recursively constructed metal artifact image, enabling the latent field to capture metal-dependent global artifact patterns. After projection completion using the INR prior, we further suppress residual artifacts using a metal-conditioned correction network, where the metal mask modulates intermediate features via adaptive instance normalization to target metal-dependent secondary artifacts while preserving anatomical structures. Experiments on the public AAPM-MAR benchmark demonstrate that MGMAR achieves state-of-the-art performance, attaining an average final score of 0.89 on 29 clinical test cases.

Abstract

Explore similar work

MGMAR: Metal-Guided Metal Artifact Reduction for X-ray Computed Tomography

Self-Auditing Residual Drifting for Pathology-Preserving Accelerated Knee MRI

Next-Acceleration-Scale Prediction for Autoregressive MRI Reconstruction