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

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "MASC: Metal-Aware Sampling and Correction via Reinforcement Learning for Accelerated MRI".

Jane: Here's the gist of the paper. The gist The proposed MASC framework is a unified reinforcement learning approach that jointly optimizes metal-aware k-space sampling and artifact correction for accelerated MRI.

Tom: Problem Formulation Active MRI acquisition is formulated as a Markov Decision Process (MDP) where the state at time step t is defined as st = (It, Mt), where It represents the current magnitude reconstruction obtained from partially sampled k-space, and Mt is a binary mask indicating acquired phase-encoding lines<ref:2602.00348#pg2> The action space consists of selecting a single unacquired phase-encoding line<ref:2602.00348#pg6> Upon selecting action at, the mask is updated as Mt+1

at: = 1, and the agent receives a reward rt based on reconstruction quality improvement<ref:2602.00348#pg8> The reward is defined as rt = α · (Q(It+1, I∗) − Q(It, I∗)) where Q measures reconstruction quality by combining Structural Similarity Index (SSIM) and Normalized Mean Squared Error (NMSE)<ref:2602.00348#pg10> MASC Framework The MASC framework comprises two main components: an artifact-aware PPO-based acquisition policy and a U-Net-based MAR network<ref:2602.00348#pg2> The Acquisition Policy utilizes an artifact-aware Proximal Policy Optimization (PPO) agent that uses a shared convolutional encoder for actor and critic branches to process the concatenated reconstruction and acquisition mask<ref:2602.00348#pg6> The MAR Network is a U-Net with four encoder and four decoder stages, employing residual learning where gψ(I) = I + rψ(I) to focus on learning artifact corrections rather than full image reconstruction<ref:2602.00348#pg8> Co-Adaptive Training The framework employs a two-stage training scheme for co-adaptive optimization<ref:2602.00348#pg10> Stage 1 involves pre-training the MAR network on paired metal-corrupted and clean images using Lpretrain = LL1 + λ · LSSIM, which is effective for training from scratch<ref:2602.00348#pg10> Stage 2 involves joint optimization where the artifact-aware PPO agent observes MAR-processed reconstructions with rewards computed as rt = α · (Q(gψ(It+1), I∗) − Q(gψ(It), I∗)) and the MAR network is fine-tuned using MSE loss, Lfinetune = ∥gψ(It) − I∗∥2<ref:2602.00348#pg10> This co-adaptation allows the policy to learn acquisition patterns complementary to the MAR network, and vice versa<ref:2602.00348#pg10> Data and Experiments The training data is constructed using a paired dataset from 200 subjects in the AutoPET CT/PET dataset, generated via physics-based metal artifact simulation<ref:2602.00348#pg10> This simulation uses a physics-based simulator to generate clean k-space without metal and metal-corrupted k-space with the virtual implant, yielding exactly matched pairs for supervised training<ref:2602.00348#pg10> Experiments are conducted on an NVIDIA RTX 5090 GPU with 32 GB memory, using an acceleration factor of 10× or 5× depending on the configuration Baselines include conventional strategies like Center-out and learned baselines such as Deep Q-Network (DQN) and Subject-Specific Double Deep Q-Network (SS-DDQN) Results and Generalization MASC substantially outperforms all baselines across every metric, including those enhanced with MAR post-processing Compared to the best two-stage baseline, MASC achieves 4.0% improvement in SSIM and 20.7% reduction in MSE; compared to the best conventional baseline without MAR, the improvements are 43.9% in SSIM and 77.3% reduction in MSE at 10× acceleration Cross-dataset evaluation on the FastMRI-based Metal Artifacts Dataset demonstrates that MASC generalizes effectively across both data distributions and implant variations, maintaining superior performance despite different hip implant geometry The results demonstrate that joint optimization outperforms sequential approaches where acquisition and reconstruction are treated independently Ablation studies confirm that end-to-end training provides measurable benefits over using a frozen pre-trained MAR network, validating the importance of joint optimization The current limitation of the work is the focus on hip implants at one anatomical region Future work will extend MASC to diverse implant materials with varying magnetic susceptibilities, different anatomical regions such as spine and knee, and multi-coil settings where parallel imaging can be combined with learned acquisition policies The paper concludes that joint optimization outperforms sequential approaches where acquisition and reconstruction are treated independently.

Title and authors: Tom: So, Jane, we're looking at this paper called "MASC: Metal-Aware Sampling and Correction via Reinforcement Learning for Accelerated MRI". It tackles the problem of using accelerated MRI when there are metal implants in the patient.

Jane: It sounds like it tries to solve two separate problems at once: how to sample k-space—that’s what you acquire—and how to fix the artifacts that those metal implants cause.

Tom: Exactly, and this paper proposes a unified reinforcement learning framework for that joint optimization. It moves beyond treating them as independent tasks, which is the core idea here.

Lu: The authors are Zhengyi Lu, Ming Lu, Chongyu Qu, Junchao Zhu, Junlin Guo, Marilyn Lionts, Yanfan Zhu, Yuechen Yang and Tianyuan Yao. They have a strong background in both representation learning and reinforcement learning for imaging tasks.

Meng: It's interesting to see a collaboration across those areas since we’re often dealing with deployment challenges where you need models that actually work on real hardware, not just clean data.

Lalam: I see they are focusing on making the acquisition policy aware of the artifacts so it can sample better k-space lines. It’s about learning which lines to pick based on what kind of artifact you expect.

Jane: So, if we boil it down simply, MASC is a system that learns the best way to acquire MRI data around metal implants while simultaneously correcting those metal artifacts during reconstruction.

Tom: Right, so it’s not just a sampling strategy anymore; it’s an intelligent acquisition strategy that knows how to work with the artifact correction network.

Lu: It's about creating a loop where the acquisition policy and the artifact correction network learn together based on the final image quality.

The paper's summary: Tom: So, what does this actually do? The paper outlines MASC as an active MRI acquisition problem framed as a Markov Decision Process, or MDP.

Jane: An MDP means the system is making a sequence of decisions—picking one unacquired phase-encoding line at a time—and the reward it gets tells it how good that decision was for the final image.

Tom: The state in this model is defined by what’s already reconstructed, which they call It, and a mask indicating which lines have been acquired, Mt.

Lu: They define the action space as selecting just one unacquired phase-encoding line to acquire next. When an action is taken, the mask updates to include that line.

Meng: That reward structure is key because it’s based on how much the reconstruction quality improves from one step to the next, specifically using a combination of Structural Similarity Index or SSIM and Normalized Mean Squared Error or NMSE.

Jane: So, they're saying that instead of just trying to get a good image once at the end, you train an agent to make incremental steps that lead toward better quality.

Tom: And they propose two main components: first, an artifact-aware PPO based acquisition policy and second, a U-Net based MAR network for artifact correction.

Lu: The PPO agent uses a shared convolutional encoder for both its actor and critic branches to process the combined information of the reconstruction and the acquisition mask.

Lalam: And they also use residual learning in that U-Net, where you learn to correct artifacts by adding a residual function, gψ(I) equals I plus rψ(I), which helps it focus on fixing artifacts instead of trying to reconstruct the whole image perfectly.

The paper's improvements: Tom: Now let’s talk about how they claim MASC is better than what’s already out there. They use a two-stage co-adaptive training scheme for this joint optimization.

Jane: Stage one involves pre-training the MAR network on paired metal-corrupted and clean images using a loss function called Lpretrain, which they define as LL1 plus lambda times LSSIM.

Lu: That’s designed to get the artifact correction network started from scratch effectively by using that SSIM term.

Meng: Stage two is where the joint optimization happens. The agent observes reconstructions processed by the MAR network and gets a reward rt based on how much quality improved compared to the previous state, It, so rt equals alpha times Q(gψ(It+one), I*) minus Q(gψ(It), I*) <ref:2602.00348#pg1>.

Tom: And during this stage, they fine-tune the MAR network using a Mean Squared Error loss called Lfinetune, which is the difference between the corrected image and the target clean image.

Jane: What this co-adaptation means for us is that the acquisition policy learns sampling patterns that are complementary to what the MAR network can fix. It’s not just a fixed strategy being applied to a fixed correction network anymore.

Lu: They show that this co-adaptation allows the policy to learn acquisition patterns that work with the MAR network, and vice versa, which is crucial because they use paired data from two hundred subjects in an AutoPET CT/PET dataset generated through physics-based metal artifact simulation.

Conclusion: Tom: So, to wrap up this paper on "MASC: Metal-Aware Sampling and Correction via Reinforcement Learning for Accelerated MRI", it’s showing that joint optimization really works here.

Jane: The results are quite substantial; compared to the best two-stage baseline they tested, MASC achieves a four point zero percent improvement in SSIM and a twenty point seven percent reduction in MSE.

Lu: When we compare it to the best conventional baseline that doesn't use MAR at all, the improvements jump to forty three point nine percent in SSIM and a seventy seven point three percent reduction in MSE when using ten times acceleration.

Meng: It’s good to see they also tested this on different data distributions, like the FastMRI-based Metal Artifacts Dataset, showing it generalizes across different implant geometries.

Lalam: And the ablation studies confirm that end-to-end training provides measurable benefits over just using a frozen pre-trained MAR network.

Tom: The main thing to remember is that joint optimization outperforms sequential approaches where acquisition and reconstruction are treated independently. That’s the big finding here.

Jane: It means for clinicians, we have a system that can acquire data faster while still getting high-quality images, even in the presence of metal implants.

Lu: The limitation they mention is that their focus is currently on hip implants at one anatomical region, and future work will extend this to other materials and different body parts.

Meng: And from a deployment standpoint, they need to figure out how to make this run efficiently on actual clinical scanners, which is the next hurdle.