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

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

💡 Research Summary

**
The paper addresses the long‑standing challenge of performing fast magnetic resonance imaging (MRI) in patients who have metallic implants. Metal implants generate severe off‑resonance fields that cause signal voids, geometric distortions, and bright streaking artifacts, while conventional accelerated MRI techniques assume artifact‑free anatomy and therefore fail when metal is present. Existing work treats metal artifact reduction (MAR) and accelerated acquisition as separate problems, leaving a gap for a unified solution.

MASC (Metal‑Aware Sampling and Correction) is introduced as a reinforcement‑learning (RL) framework that jointly learns a metal‑aware k‑space sampling policy and a MAR reconstruction network. The key idea is to treat active k‑space acquisition as a Markov decision process (MDP) where, at each time step, the agent observes the current partially reconstructed image (obtained by inverse Fourier transform of the already sampled k‑space) together with a binary mask indicating which phase‑encoding lines have been acquired. The action space consists of selecting one previously unsampled line. After the action, the mask is updated, a new partial reconstruction is computed, and a reward is issued. The reward is defined as the improvement in a quality metric Q between the new and previous reconstructions, where Q is a weighted sum of Structural Similarity Index (SSIM) and Normalized Mean Squared Error (NMSE). This formulation directly encourages the agent to acquire those k‑space lines that most reduce metal‑induced artifacts.

The reconstruction network is a U‑Net with residual learning (the network predicts a correction term that is added to the input). In the first training stage the MAR network is pre‑trained on paired metal‑corrupted and clean images using a combination of L1 loss and SSIM loss. This provides a strong baseline that can already suppress metal artifacts when the full k‑space is available.

The second stage performs joint optimization. The PPO (Proximal Policy Optimization) agent receives as input the MAR‑corrected partial reconstructions, and the MAR network is fine‑tuned with an MSE loss on the same partial reconstructions. The two components are updated alternately after each rollout, allowing them to co‑adapt: the MAR network learns to handle the specific aliasing patterns produced by the learned sampling policy, while the policy learns to select k‑space lines that best support the MAR network’s correction capability.

A crucial contribution is the creation of a paired MRI dataset that enables supervised learning for both MAR and RL reward computation. Starting from 200 subjects in the AutoPET CT dataset, the authors use TotalSegmentator to generate multi‑tissue 3‑D phantoms (including proton density, T1, T2 maps). A virtual cobalt‑chromium hip prosthesis is manually placed in the hip region of each phantom. Using an open‑source physics‑based MRI simulator (Zóchowski et al., 2024), they generate full‑resolution TSE (Turbo Spin Echo) k‑space data at 3 T both with and without the implant, as well as a binary implant mask. The result is a set of exactly matched clean‑metal image pairs for 200 subjects, each with 36 slices, providing ground‑truth for MAR training and a reliable signal for reward calculation.

For cross‑dataset validation, a second test set is built from the FastMRI knee dataset. Metal artifacts are synthetically added using a different simulation pipeline (different implant geometry, RF profile, and off‑resonance model) to assess generalization.

Experimental results compare MASC against several baselines: uniform undersampling, variable‑density sampling, and two recent RL‑based acquisition methods that were originally designed for artifact‑free data. Across all metrics (SSIM, NMSE, PSNR) MASC achieves the highest scores, with SSIM ≈ 0.92 and NMSE ≈ 0.03 at a 30 % sampling budget. Visual inspection shows a marked reduction of streaking and signal loss around the metal, confirming that the learned policy focuses on k‑space regions most affected by susceptibility‑induced field perturbations. The joint training scheme further improves performance over a “frozen” MAR network, demonstrating the benefit of co‑adaptation.

Inference speed is practical: with both the PPO policy and MAR U‑Net frozen, the full pipeline processes a slice in under one second on a single GPU, comparable to conventional accelerated MRI pipelines and suitable for clinical use.

Limitations include reliance on simulated data (real patient scans with actual implants were not used), the current 2‑D slice‑wise formulation (3‑D temporal consistency is not modeled), and the need to evaluate on a broader variety of implant materials and locations. Future work should extend the RL policy to 3‑D acquisition, incorporate real‑world metal‑implant datasets, and explore transfer learning to other imaging modalities.

In summary, MASC is the first framework that simultaneously learns metal‑aware k‑space sampling and artifact correction via reinforcement learning, leverages a novel paired metal‑clean MRI dataset, and demonstrates superior reconstruction quality and generalization on both simulated hip and knee data. It opens a promising direction for integrating physics‑based simulation, deep learning, and RL to tackle the dual challenges of artifact reduction and scan acceleration in MRI.