MRI Reconstruction with Interpretable Pixel-Wise Operations Using Reinforcement Learning

Wentian Li, Xidong Feng, Haotian An, Xiang Yao Ng, Yu-Jin ZhangAAAI 2020 · DOI

Introduction

Magnetic Resonance Imaging (MRI) is a critical diagnostic tool, yet its utility is often limited by slow acquisition times. To accelerate scans, technicians frequently perform "down-sampling," skipping portions of the frequency domain (k-space). While this reduces scan time, it introduces significant blurring and aliasing artifacts.

Modern deep learning solutions, such as Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), have shown remarkable success in removing these artifacts. However, these models typically function as "black boxes," offering little insight into their decision-making process. In clinical settings, this lack of transparency is a liability; if a model erroneously smooths a lesion or introduces a hallucinated structure, clinicians have no way to audit the process. This paper introduces a novel reinforcement learning (RL) framework that replaces opaque architectures with a sequence of transparent, interpretable pixel-wise operations.

Methodology

The authors formulate MRI reconstruction as a Markov Decision Process (MDP), where every pixel in the image is treated as an autonomous agent. Instead of a single network processing the entire image, these agents learn to apply a sequence of local filters to recover the underlying anatomy.

Hybrid RL Architecture

The system utilizes a specialized architecture to manage both the selection and the magnitude of image processing operations:

  • Discrete Policy: An Advantage Actor-Critic (A2C) framework determines which of the 12 available filters (e.g., box, bilateral, Sobel, or "do nothing") to apply at each step.
  • Continuous Parameters: For filters requiring specific intensity settings (like unsharp masking or Laplace filters), a modified Deep Deterministic Policy Gradient (DDPG) learns the optimal strength.
  • Data Consistency (DC): Following the RL-based filtering (typically 3 steps), a final Data Consistency layer ensures the reconstructed image remains faithful to the original raw k-space data.
Hybrid RL architecture for pixel-wise reconstruction, showing the shared feature extractor and specialized heads for value, parameters, and policy.

Experimental Setup

The model was evaluated on two major benchmarks:

  1. MICCAI 2013 Grand Challenge: T1-weighted brain tissue images (16,095 training samples).
  2. fastMRI (NYU Langone Health): A large-scale knee imaging dataset, focusing on the single-coil subset.

Key Insights and Results

Anatomical Alignment and Learned Pipelines

A striking outcome of this approach is that agents "learn" anatomy. The model automatically applies edge-enhancing Sobel filters along complex contours like the skull and brain sulcus, while opting for smoothing filters in uniform tissue regions. Furthermore, the agents spontaneously discovered effective image processing pipelines—such as Unsharp Masking → Laplace Filter → Bilateral/Median Filter—mimicking the manual tuning performed by human experts.

Quantitative and Qualitative Performance

The RL method achieved state-of-the-art results, performing favorably against both classical optimization and modern deep learning baselines:

  • MICCAI Results: At a 20% sampling rate, the method achieved a PSNR of 39.52, outperforming DAGAN (39.44) and BM3D (39.40).
  • fastMRI Results: On the knee dataset, the model matched standard U-Net performance with an NMSE of 0.14.
Reconstruction results on brain (top) and knee (bottom) datasets. The figure compares the target image, the undersampling mask, zero-filled results, and the proposed method with and without the Data Consistency (DC) step.

Qualitatively, the RL approach avoids common pitfalls of GAN-based methods, such as "checkerboard" artifacts or artificial textures, producing smoother and more natural images in flat tissue regions.

Conclusion

By treating pixels as agents with an interpretable action space, this framework bridges the gap between high-performance deep learning and clinical interpretability. While the current model relies on local operations and may struggle with large-scale structural artifacts at extremely low sampling rates (e.g., 10%), it provides a transparent audit trail for every pixel. Future work will explore operations that consider longer-range pixel relationships and integrate these interpretable "front-ends" with classical reconstruction algorithms for enhanced stability.

Summary updated

BibTeX
@inproceedings{li2020mri,
  author = {Wentian Li and Xidong Feng and Haotian An and Xiang Yao Ng and Yu-Jin Zhang},
  title = {{MRI Reconstruction with Interpretable Pixel-Wise Operations Using Reinforcement Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i01.5423},
}
Figure