InDuDoNet+: A Model-Driven Interpretable Dual Domain Network for Metal Artifact Reduction in CT Images

Hong Wang, Yuexiang Li, Haimiao Zhang, Deyu Meng, Yefeng Zheng

Abstract

During the computed tomography (CT) imaging process, metallic implants within patients often cause harmful artifacts, which adversely degrade the visual quality of reconstructed CT images and negatively affect the subsequent clinical diagnosis. For the metal artifact reduction (MAR) task, current deep learning based methods have achieved promising performance. However, most of them share two main common limitations: 1) the CT physical imaging geometry constraint is not comprehensively incorporated into deep network structures; 2) the entire framework has weak interpretability for the specific MAR task; hence, the role of each network module is difficult to be evaluated. To alleviate these issues, in the paper, we construct a novel deep unfolding dual domain network, termed InDuDoNet+, into which CT imaging process is finely embedded. Concretely, we derive a joint spatial and Radon domain reconstruction model and propose an optimization algorithm with only simple operators for solving it. By unfolding the iterative steps involved in the proposed algorithm into the corresponding network modules, we easily build the InDuDoNet+ with clear interpretability. Furthermore, we analyze the CT values among different tissues, and merge the prior observations into a prior network for our InDuDoNet+, which significantly improve its generalization performance. Comprehensive experiments on synthesized data and clinical data substantiate the superiority of the proposed methods as well as the superior generalization performance beyond the current state-of-the-art (SOTA) MAR methods.

Introduction: The Challenge of Metal Artifacts

Metallic implants, such as hip replacements or dental fillings, present a significant challenge in Computed Tomography (CT) imaging. Because metal absorbs X-rays much more strongly than human tissue, it creates "missing" data in the raw projection data, known as a sinogram. When this incomplete data is reconstructed into a 3D image, it manifests as severe dark shadings and bright streaks. These artifacts obscure critical anatomy and can make clinical diagnosis nearly impossible.

While deep learning has shown promise in addressing this, existing methods often suffer from two major flaws: they act as "black boxes" that ignore the physical geometry of CT scanners, and they lack interpretability, making it difficult for clinicians to trust the modifications made to medical images.

Methodology: Interpretable Dual Domain Unfolding

The proposed InDuDoNet+ (Interpretable Dual Domain Network) addresses these limitations by embedding the CT imaging process directly into the network architecture. The framework is built upon the concept of Deep Unfolding, where a traditional iterative optimization algorithm is translated into a neural network. Each iteration of the mathematical model becomes one "stage" in the network, ensuring the system follows a logical, physics-based path to the solution.

Architecture of InDuDoNet+ showing the entire network (a) and the internal structure of the n-th stage (b), highlighting the dual-domain interaction between S-net and X-net.

Key technical innovations include:

  • Dual Domain Processing: The network operates simultaneously in the Spatial Domain (the CT image) and the Radon Domain (the sinogram). Information is passed back and forth between these domains using the physical geometry of the CT scanner (forward and back projections).
  • Knowledge-Driven Prior-net: This module uses a lightweight sub-network (WNet) to normalize the sinogram based on known CT values for different tissues (air, soft tissue, bone). By making the data more homogeneous before main processing, the network can focus specifically on difficult metal artifacts rather than basic anatomy.

Experimental Evaluation

The researchers evaluated InDuDoNet+ across three distinct scenarios to test accuracy and robustness:

  1. Synthesized DeepLesion: A large-scale dataset using 1,000 clean CT images with 100 different simulated metallic implant shapes.
  2. Synthesized Dental: A test of generalization where the model, trained on body scans (thorax/abdomen), was tested on dental scans with entirely different anatomical structures.
  3. Clinical SpineWeb: Real-world clinical data containing actual metal artifacts to validate the method's performance outside of simulation.

Key Results and Performance

InDuDoNet+ demonstrates significant improvements over existing state-of-the-art (SOTA) methods in both quantitative metrics and visual quality. On the DeepLesion dataset, it achieved a Peak Signal-to-Noise Ratio (PSNR) of 41.50 dB, significantly outperforming previous models like DuDoNet++ (39.69 dB).

Example of a CT image processed by InDuDoNet, demonstrating the reduction of metal artifacts and preservation of anatomical detail.

Beyond accuracy, the model is highly efficient. By following a specific mathematical optimization, it requires only 1.78 million parameters—a fraction of the 25 million parameters used by competing models. Furthermore, its processing speed of approximately 0.37 seconds per image makes it viable for clinical workflows. The "cross-body-site" test also confirmed superior generalization, as the model successfully adapted to dental images despite being trained on thorax data.

Conclusion and Future Directions

By treating physics as a regularizer, InDuDoNet+ prevents the "hallucinations" common in generic AI models. However, some limitations remain. The current model relies on thresholding for metal segmentation and requires supervised learning with "ground truth" clean images. Future work will focus on automatic metal localization and exploring semi-supervised or unsupervised learning to better handle real-world clinical data where perfect, artifact-free scans are unavailable for training.

Summary updated

Abstract

During the computed tomography (CT) imaging process, metallic implants within patients often cause harmful artifacts, which adversely degrade the visual quality of reconstructed CT images and negatively affect the subsequent clinical diagnosis. For the metal artifact reduction (MAR) task, current deep learning based methods have achieved promising performance. However, most of them share two main common limitations: 1) the CT physical imaging geometry constraint is not comprehensively incorporated into deep network structures; 2) the entire framework has weak interpretability for the specific MAR task; hence, the role of each network module is difficult to be evaluated. To alleviate these issues, in the paper, we construct a novel deep unfolding dual domain network, termed InDuDoNet+, into which CT imaging process is finely embedded. Concretely, we derive a joint spatial and Radon domain reconstruction model and propose an optimization algorithm with only simple operators for solving it. By unfolding the iterative steps involved in the proposed algorithm into the corresponding network modules, we easily build the InDuDoNet+ with clear interpretability. Furthermore, we analyze the CT values among different tissues, and merge the prior observations into a prior network for our InDuDoNet+, which significantly improve its generalization performance.Comprehensive experiments on synthesized data and clinical data substantiate the superiority of the proposed methods as well as the superior generalization performance beyond the current state-of-the-art (SOTA) MAR methods. . Code is available at \url{https://github.com/hongwang01/InDuDoNet_plus}.
BibTeX
@article{wang2023indudonet,
  author = {Hong Wang and Yuexiang Li and Haimiao Zhang and Deyu Meng and Yefeng Zheng},
  title = {{InDuDoNet+: A Model-Driven Interpretable Dual Domain Network for Metal Artifact Reduction in CT Images}},
  journal = {Medical Image Analysis},
  year = {2023},
  doi = {10.1016/j.media.2022.102729},
}
Figure