MASC: Metal-Aware Sampling and Correction via Reinforcement Learning for Accelerated MRI

TL;DR

This work tackles the problem of severe metal-induced artifacts in MRI while also pursuing accelerated acquisition. It introduces MASC, a unified framework that jointly optimizes an artifact-aware sampling policy via Proximal Policy Optimization and a MAR U-Net, trained end-to-end with a physics-based paired dataset that provides exactly matched clean and metal-affected references. Key contributions include a co-adaptive training scheme, a physics-based MRI dataset for supervised MAR and reward computation, and cross-dataset validation demonstrating generalization to realistic clinical data. The results show that joint optimization yields substantial improvements over conventional and two-stage methods, suggesting strong potential for clinically robust MAR with accelerated MRI.

Abstract

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 (5)

• Figure 1: Overview of MASC.Top: Existing approaches using conventional or learning-based sampling strategies are trained on metal-free data and produce clean results when tested on implant-free cases, but fail when applied to metal-corrupted data, yielding artifact-unaware reconstructions. Bottom: MASC jointly optimizes a MAR U-Net and active sampling policy through end-to-end training on metal-corrupted data, enabling artifact-aware k-space sampling and effective artifact reduction.
• Figure 2: MASC training and inference pipeline.Top (Train): Starting from full metal-corrupted k-space, the current mask selects acquired lines for IFFT reconstruction. The MAR U-Net processes the reconstruction to produce artifact-reduced images, which the artifact-aware PPO agent observes to determine the next sampling action. Color-coded arrows indicate operation frequency: purple for pre-training (MAR U-Net on paired metal-clean images), blue for per-step operations (sequential k-space acquisition), and orange for per-rollout updates (reward computation, PPO policy update, and MAR U-Net fine-tuning). Fire icons denote trainable networks. Bottom (Inference): Both networks are frozen (snowflake icons). The agent iteratively selects k-space lines until the acquisition budget is exhausted.
• Figure 3: Dataset construction pipeline.Phantom generation: CT volumes from the autoPET dataset are processed through TotalSegmentator with three complementary tasks to produce multi-tissue phantom maps. The hip region is manually selected and combined with a cobalt-chromium hip implant model to generate tissue property maps including implant mask, off-resonance frequency (df), susceptibility (susp), proton density (PD), T1, and T2. MRI simulation: A physics-based simulator with TSE sequence parameters generates paired k-space data and reconstructions: clean images without metal and artifact-corrupted images with the virtual implant, providing exactly matched pairs for supervised training.
• Figure 4: Comparison of acquisition strategies.Top: Ground truth, sampling masks, reconstructions, and error maps for each method including four conventional baselines (Random, Random Low-biased, Equispaced, Center-out), learned baseline (DQN, SS-DDQN), and our MASC. MASC produces reconstructions closest to ground truth with substantially darker error maps. Bottom: SSIM and MSE versus number of acquired k-space lines. Shaded regions indicate the std. MASC demonstrates superior performance throughout acquisition, with the gap widening as more lines are acquired.
• Figure 5: Ablation study on training strategy components at $10\times$ acceleration. Configurations progressively add metal-aware training, MAR integration, and end-to-end optimization. SSIM, MSE, PSNR and MAE versus number of acquired k-space lines. Shaded regions indicate the std. MASC demonstrates superior performance throughout acquisition, with the gap widening as more lines are acquired.