Artifact Disentanglement Network for Unsupervised Metal Artifact Reduction (1906.01806v5)

Abstract: Current deep neural network based approaches to computed tomography (CT) metal artifact reduction (MAR) are supervised methods which rely heavily on synthesized data for training. However, as synthesized data may not perfectly simulate the underlying physical mechanisms of CT imaging, the supervised methods often generalize poorly to clinical applications. To address this problem, we propose, to the best of our knowledge, the first unsupervised learning approach to MAR. Specifically, we introduce a novel artifact disentanglement network that enables different forms of generations and regularizations between the artifact-affected and artifact-free image domains to support unsupervised learning. Extensive experiments show that our method significantly outperforms the existing unsupervised models for image-to-image translation problems, and achieves comparable performance to existing supervised models on a synthesized dataset. When applied to clinical datasets, our method achieves considerable improvements over the supervised models. The source code of this paper is publicly available at https://github.com/liaohaofu/adn.

• The paper introduces an unsupervised framework that separates metal artifacts from anatomical features using adversarial and self-regularization techniques.
• The ADN employs innovative encoders and multiple loss functions to achieve competitive PSNR and SSIM results compared to state-of-the-art supervised models.
• The method demonstrates robust performance in clinical settings, effectively handling unpaired CT and CBCT data for enhanced diagnostic imaging.

Artifact Disentanglement Network for Unsupervised Metal Artifact Reduction: An Expert Overview

The paper "Artifact Disentanglement Network for Unsupervised Metal Artifact Reduction" introduces a novel approach to metal artifact reduction (MAR) in computed tomography (CT) imaging. The paper presents an Artifact Disentanglement Network (ADN) as the first unsupervised machine learning framework designed specifically for MAR. This research is particularly timely given the limitations of existing supervised methods, which heavily rely on synthetic data that may not accurately reflect clinical scenarios, often resulting in reduced performance in real-world applications due to significant domain shifts.

Contributions and Methodology

This paper makes a critical contribution by addressing the challenge of MAR in an unsupervised setting, eliminating the need for paired CT images during training. The authors accomplish this by leveraging an innovative artifact disentanglement framework that separates metal artifacts from anatomical content in the latent space of images. This disentanglement allows for versatile manipulations and transformations between artifact-affected and artifact-free images, achieving various adversarial and self-regularizations.

The proposed ADN comprises several key components:

• Artifact-Free Image Encoder, Generator, and Discriminator: These elements work together to map artifact-free content from images into a common latent space, allowing for effective adversarial learning.
• Artifact-Affected Image Encoder and Discriminator: These components focus on identifying and processing images with metal artifacts, reinforcing the network's ability to distinguish between different artifact states.
• Artifact-Only Encoder: This encoder isolates the artifact-specific information, supporting the disentanglement and synthesis of artifacts within the learning framework.

The paper employs a series of loss functions to ensure both MAR efficacy and anatomical content preservation. These include adversarial, self-reconstruction, cycle-consistency, and artifact-consistent losses, which collectively guide the network towards generating artifact-corrected outputs without the need for paired data.

Experimental Evaluation

The empirical evaluation presented in the paper covers a synthesized dataset (SYN) and two clinical datasets (CL1 and CL2), encompassing both CT and cone-beam CT (CBCT) modalities. On the synthetic benchmark, the unsupervised ADN demonstrates comparable performance to state-of-the-art supervised methods, including CNNMAR and cGANMAR, in terms of peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Notably, the ADN surpasses traditional unsupervised frameworks like CycleGAN and MUNIT, particularly in qualitative assessments, by effectively reducing artifacts and preserving anatomical structures.

The performance on clinical datasets is especially compelling, with the ADN outperforming supervised models that struggle with domain adaptation. In scenarios involving unpaired clinical data (CL1) and cross-modality settings (CL2), the ADN demonstrates superior robustness and adaptability, reinforcing its practical applicability in real-world medical imaging environments.

Implications and Future Directions

The outcome of this research has significant implications for the field of medical imaging, particularly regarding the deployment of deep learning models in clinical settings. By circumventing the dependency on synthetic data and harnessing unsupervised learning strategies, the proposed method offers a more feasible and reliable alternative for MAR, potentially augmenting imaging quality and diagnostic efficacy in practice.

Looking forward, this work opens new avenues for unsupervised learning applications in medical image analysis. Future developments may involve extending the ADN architecture to other types of imaging artifacts, exploring more efficient network designs for real-time applications, and further refining the adversarial training schemes to improve convergence and stability. The interplay between artifact-specific encoders and multi-scale generative frameworks also warrants deeper investigation, as it could enhance the network's capability in handling complex imaging scenarios.

In conclusion, this paper presents a well-conceived and empirically validated framework for unsupervised MAR, offering valuable insights and a practical solution to a longstanding challenge in CT imaging. The authors have set a strong foundation for future innovations in unsupervised medical image processing.