MASC: Metal-Aware Sampling and Correction via Reinforcement Learning for Accelerated MRI
Authors: Zhengyi Lu, Ming Lu, Chongyu Qu, Junchao Zhu, Junlin Guo, Marilyn Lionts, Yanfan Zhu, Yuechen Yang, Tianyuan Yao, Jayasai Rajagopal, Bennett Allan Landman, Xiao Wang, Xinqiang Yan, Yuankai Huo
Organizations: Vanderbilt University, Nashville, TN, USA 37215 · Vanderbilt University Medical Center, Nashville, TN, USA 37232 · Oak Ridge National Laboratory, Oak Ridge, TN, USA 37831
Metal implants in MRI cause severe artifacts that degrade image quality and hinder clinical diagnosis. Traditional approaches address metal artifact reduction (MAR) and accelerated MRI acquisition as separate problems. We propose MASC, a unified reinforcement learning framework that jointly optimizes metal-aware k-space sampling and artifact correction for accelerated MRI. To enable supervised training, we construct a paired MRI dataset using physics-based simulation, generating k-space data and reconstructions for phantoms with and without metal implants. This paired dataset provides simulated 3D MRI scans with and without metal implants, where each metal-corrupted sample has an exactly matched clean reference, enabling direct supervision for both artifact reduction and acquisition policy learning. We formulate active MRI acquisition as a sequential decision-making problem, where an artifact-aware Proximal Policy Optimization (PPO) agent learns to select k-space phase-encoding lines under a limited acquisition budget. The agent operates on undersampled reconstructions processed through a U-Net-based MAR network, learning patterns that maximize reconstruction quality. We further propose an end-to-end training scheme where the acquisition policy learns to select k-space lines that best support artifact removal while the MAR network simultaneously adapts to the resulting undersampling patterns. Experiments demonstrate that MASC's learned policies outperform conventional sampling strategies, and end-to-end training improves performance compared to using a frozen pre-trained MAR network, validating the benefit of joint optimization. Cross-dataset experiments on FastMRI with physics-based artifact simulation further confirm generalization to realistic clinical MRI data. The code and models of MASC have been made publicly available: https://github.com/hrlblab/masc
Figures & tables
Figure 1Figure 2Figure 3Figure 4
Method
MAR
SSIM ↑
PSNR ↑
MSE ↓
NMSE ↓
MAE ↓
Random
×
0.3214 ± 0.0440
11.72 ± 1.11
0.0695 ± 0.0173
0.3374 ± 0.0929
0.2199 ± 0.0333
✓
0.6949 ± 0.0463
18.79 ± 1.42
0.0140 ± 0.0049
0.0680 ± 0.0243
0.0708 ± 0.0143
Random-LowBias
×
0.3556 ± 0.0444
12.70 ± 1.38
0.0566 ± 0.0193
0.2696 ± 0.0718
0.1877 ± 0.0332
✓
0.6929 ± 0.0516
19.10 ± 1.45
0.0131 ± 0.0056
0.0635 ± 0.0241
0.0716 ± 0.0170
Equispaced
×
0.2648 ± 0.0380
11.26 ± 0.92
0.0764 ± 0.0152
0.3716 ± 0.0846
0.2331 ± 0.0297
✓
0.4085 ± 0.0478
12.82 ± 1.05
0.0538 ± 0.0131
0.2607 ± 0.0659
0.1546 ± 0.0237
Table 1: Comparison with baseline acquisition strategies at 10× acceleration (2 initial + 18 budget lines). Baselines are evaluated with and without MAR. Results are mean ± std on the test set. Best results in bold . All improvements are statistically significant ( p<0.001 , paired t-test).
Configuration
SSIM ↑
PSNR ↑
MSE ↓
NMSE ↓
MAE ↓
Clean-trained PPO
0.5120 ± 0.0559
14.47 ± 3.19
0.0463 ± 0.0348
0.2137 ± 0.1309
0.1465 ± 0.0606
Metal-trained PPO
0.4758 ± 0.0465
14.99 ± 2.47
0.0373 ± 0.0220
0.1766 ± 0.0926
0.1341 ± 0.0425
Clean PPO + Pretrained MAR
0.6640 ± 0.0600
17.08 ± 1.68
0.0212 ± 0.0095
0.1020 ± 0.0400
0.0945 ± 0.0262
Metal PPO + Pretrained MAR
0.6874 ± 0.0621
17.72 ± 1.59
0.0182 ± 0.0077
0.0888 ± 0.0384
0.0820 ± 0.0213
MASC (Ours)
0.7224 ± 0.0408
19.73 ± 1.28
0.0111 ± 0.0036
0.0546 ± 0.0190
0.0640 ± 0.0121
Table 2: Ablation study on training strategy components at 10× acceleration. Configurations progressively add metal-aware training, MAR integration, and end-to-end optimization. Results are mean ± std on the test set. Best results in bold . MASC significantly outperforms all ablation variants ( p<0.001 , paired t-test).
Method (with MAR)
SSIM ↑
PSNR ↑
MSE ↓
NMSE ↓
MAE ↓
Random
0.5562 ± 0.0434
20.41 ± 2.73
0.0111 ± 0.0079
0.0642 ± 0.0762
0.0773 ± 0.0282
Random-LowBias
0.5636 ± 0.0428
20.41 ± 2.56
0.0108 ± 0.0064
0.0558 ± 0.0305
0.0767 ± 0.0245
Equispaced
0.5402 ± 0.0465
20.12 ± 2.57
0.0116 ± 0.0071
0.0663 ± 0.0652
0.0790 ± 0.0262
Center-out
0.5072 ± 0.0552
17.79 ± 2.67
0.0199 ± 0.0119
0.0994 ± 0.0417
0.1096 ± 0.0373
DQN
0.5376 ± 0.0483
19.52 ± 2.36
0.0129 ± 0.0068
0.0686 ± 0.0375
0.0851 ± 0.0255
SS-DDQN
0.5305 ± 0.0487
19.55 ± 2.36
0.0127 ± 0.0066
0.0682 ± 0.0389
0.0842 ± 0.0244
Table 3: Cross-dataset generalization at 10× acceleration. All methods are trained on our physics-based simulation dataset and evaluated on the FastMRI-based Metal Artifacts Dataset. The baselines are post-processed with MAR for fair comparison. Results are mean ± std on the test set. Best results in bold . MASC significantly outperforms all baselines ( p<0.001 , paired t-test).
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Characteristic
Value
Subjects
200
Age (years)
62.0 ± 15.4 (range: 11–95)
Sex
Male
115 (57.5%)
Female
84 (42.0%)
Not reported
1 (0.5%)
Appendix
Table 4: Dataset demographics (n=200). Summary of patient characteristics from the AutoPET dataset used for MRI simulation.
Method
MAR
SSIM ↑
PSNR ↑
MSE ↓
NMSE ↓
MAE ↓
Random
×
0.4037 ± 0.0351
13.95 ± 2.37
0.0471 ± 0.0290
0.2193 ± 0.1143
0.1518 ± 0.0453
✓
0.6789 ± 0.0683
18.36 ± 1.87
0.0162 ± 0.0089
0.0760 ± 0.0358
0.0861 ± 0.0262
Random-LowBias
×
0.4317 ± 0.0441
13.59 ± 2.94
0.0544 ± 0.0361
0.2525 ± 0.1441
0.1616 ± 0.0562
✓
0.6580 ± 0.0875
18.39 ± 2.39
0.0175 ± 0.0156
0.0809 ± 0.0579
0.0885 ± 0.0371
Equispaced
×
0.3670 ± 0.0252
13.77 ± 2.26
0.0486 ± 0.0288
0.2254 ± 0.1120
0.1555 ± 0.0443
✓
0.5716 ± 0.0585
17.01 ± 1.56
0.0214 ± 0.0099
0.1018 ± 0.0391
0.0980 ± 0.0256
Appendix
Table 5: Comparison with baseline acquisition strategies at 5× acceleration (8 initial + 32 budget lines). Baselines are evaluated with and without metal artifact reduction (MAR). Results are mean ± std on the test set. Best results in bold . MASC significantly outperforms all baselines ( p<0.001 , paired t-test).
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× and 8× acceleration demonstrate consistent improvements over conventional and learned baselines on both trajectories.
Zhengyi Lu, Ming Lu, Chongyu Qu +11
Vanderbilt University, Nashville, TN, USA · Vanderbilt University Medical Center, Nashville, TN, USA · Oak Ridge National Laboratory, Oak Ridge, TN, USA
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.
Hyoung Suk Park, Kiwan Jeon
National Institute for Mathematical Sciences, Daejeon, 34047, Republic of Korea
MRI reconstruction is an inherently ill-posed inverse problem, since incomplete measurements admit many plausible solutions. This ambiguity becomes more severe under high acceleration, where pixel-domain continuous predictors tend to average over feasible reconstructions and suppress high-frequency anatomy. We address this limitation by moving reconstruction to discrete multi-scale latent space and posing it as autoregressive next-acceleration-scale prediction. Leveraging discrete priors proven effective in visual autoregressive modeling, our method restricts the solution to compact sequences of codebook tokens, enabling sharp reconstructions even from extremely sparse measurements. This discrete autoregressive formulation also aligns naturally with modern large language model post-training techniques. Building on this observation, we introduce on-policy privileged information distillation for visual autoregressive modeling, where a teacher is provided training only privileged context that is unavailable at inference, in our case fully sampled acquisitions, and supervises a student trained on its own rollouts, leading to consistent reconstruction gains. Through extensive experiments on the fastMRI benchmark, we show that our approach delivers improved reconstruction performance across diverse sampling patterns under extreme undersampling. Project website is here.