MEPNet: A Model-Driven Equivariant Proximal Network for Joint Sparse-View Reconstruction and Metal Artifact Reduction in CT Images

Hong Wang, Minghao Zhou, Dong Wei, Yuexiang Li, Yefeng Zheng

Introduction

Computed Tomography (CT) is a cornerstone of modern medical diagnostics, yet it faces persistent challenges related to image quality and patient safety. To minimize radiation exposure and accelerate scan times, clinicians often utilize sparse-view CT, which involves taking fewer projections around the patient. However, this reduction in data typically introduces significant streaking artifacts. The problem is further compounded when patients have metallic implants, such as hip replacements or dental fillings. These high-density objects cause "photon starvation" and beam hardening, resulting in severe dark and bright streaks that can obscure critical anatomical structures.

While deep learning has significantly advanced CT reconstruction, many existing solutions treat sparse-view and metal artifacts as isolated issues or rely on "black-box" architectures like U-Nets. These standard models often fail to incorporate the underlying physics of CT geometry and do not account for the inherent rotational symmetry of medical anatomy. This paper introduces MEPNet (Model-Driven Equivariant Proximal Network), a framework designed to address these dual challenges simultaneously by integrating physical principles directly into the neural network architecture.

Methodology

MEPNet distinguishes itself through two primary design innovations: Model-Driven Unrolling and Rotation-Equivariant Convolutional Neural Networks (CNNs).

Model-Driven Unrolling

Rather than employing a generic deep learning architecture, the researchers derived MEPNet from a formal mathematical optimization problem. This problem balances data consistency—ensuring the reconstruction matches the raw "sinogram" data—with image priors that define what a clean CT scan should look like. By "unrolling" the iterative steps of the Proximal Gradient technique into a series of neural network layers, the model's internal logic is forced to align with the physical process of CT reconstruction. This structured approach reduces the likelihood of the model "hallucinating" features that are not present in the raw data.

Rotation-Equivariant CNNs

A key innovation of MEPNet is the use of rotation-equivariant layers. Standard CNNs are translation-equivariant (recognizing a feature regardless of its location), but they are not naturally equipped to handle rotations. In CT imaging, the same anatomical structures are viewed from multiple angles as the scanner rotates. MEPNet utilizes specialized convolutional layers based on Fourier series expansion to ensure that the network treats rotated versions of the same anatomy consistently. This embeds a powerful physical prior: the laws of physics and the nature of human anatomy do not change simply because the scanner rotates. This symmetry allows the model to be more parameter-efficient, reducing the parameter count from approximately 5.1 million in baseline models to 4.7 million.

Experimental Evaluation

The researchers evaluated MEPNet across three distinct datasets to test its performance, robustness, and ability to generalize to unseen data:

  • DeepLesion: A large-scale public dataset used for primary training and testing (2000 test pairs).
  • Pancreas-test: 50 pairs used to evaluate "cross-domain" performance, where the model is tested on a body part it did not see during training.
  • CLINIC-test: 3,397 slices from a separate clinical source to ensure real-world robustness.

Artifacts were synthesized by simulating 8x, 4x, and 2x under-sampling rates and overlaying various metal shapes to mimic clinical scenarios.

Results and Discussion

MEPNet consistently outperformed existing state-of-the-art methods, such as DuDoNet and InDuDoNet, across both quantitative and qualitative metrics.

Quantitative Performance

On the DeepLesion test set with 4x under-sampling, MEPNet achieved a Peak Signal-to-Noise Ratio (PSNR) of 41.43 dB, significantly higher than the 40.24 dB achieved by the next best method. Furthermore, it reached a Structural Similarity Index (SSIM) of 0.9889, indicating a very high degree of anatomical fidelity.

Generalization and Efficiency

One of the most significant findings was MEPNet's ability to generalize. Because of the rotational prior, the model performed remarkably well on the Pancreas-test set, despite being trained on different anatomical regions. This suggests that the model is learning fundamental reconstruction principles rather than just memorizing specific training orientations.

Visual Quality

Qualitative analysis confirms that MEPNet effectively suppresses the "star-pattern" streaks associated with metal implants and the aliasing streaks caused by sparse-view sampling. Unlike other models that may blur edges while removing noise, MEPNet preserves the sharp boundaries of bones and soft tissues.

Conclusion

MEPNet represents a significant step forward in joint CT reconstruction and artifact reduction. By combining the rigor of mathematical optimization with the flexibility of rotation-equivariant deep learning, the framework achieves superior image quality with fewer parameters. While the current study relies on simulated data, the results demonstrate a robust potential for clinical application. Future work will focus on bridging the "domain gap" by validating the model on real-world clinical scans containing physical metal implants and varying noise profiles.

Summary updated

BibTeX
@inproceedings{wang2023mepnet,
  author = {Hong Wang and Minghao Zhou and Dong Wei and Yuexiang Li and Yefeng Zheng},
  title = {{MEPNet: A Model-Driven Equivariant Proximal Network for Joint Sparse-View Reconstruction and Metal Artifact Reduction in CT Images}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2023},
  url = {https://openreview.net/forum?id=9IfpiS-mbN},
}
Figure