ADN: Artifact Disentanglement Network for Unsupervised Metal Artifact Reduction (1908.01104v4)

Abstract: Current deep neural network based approaches to computed tomography (CT) metal artifact reduction (MAR) are supervised methods that rely on synthesized metal artifacts for training. However, as synthesized data may not accurately 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 disentangles the metal artifacts from CT images in the latent space. It supports different forms of generations (artifact reduction, artifact transfer, and self-reconstruction, etc.) with specialized loss functions to obviate the need for supervision with synthesized data. Extensive experiments show that when applied to a synthesized dataset, our method addresses metal artifacts significantly better than the existing unsupervised models designed for natural image-to-image translation problems, and achieves comparable performance to existing supervised models for MAR. When applied to clinical datasets, our method demonstrates better generalization ability over the supervised models. The source code of this paper is publicly available at https://github.com/liaohaofu/adn.

• The paper introduces ADN, an unsupervised framework that disentangles CT image content and metal artifacts for robust metal artifact reduction.
• ADN employs adversarial learning and tailored losses to reconstruct clean images while preserving anatomical structure.
• Evaluations on synthesized and clinical datasets show ADN achieves competitive PSNR and SSIM, demonstrating real-world applicability.

Artifact Disentanglement Network for Unsupervised Metal Artifact Reduction

The paper introduces a novel approach in the domain of Computed Tomography (CT) imaging, particularly addressing the persistent issue of metal artifact reduction (MAR). Previous methodologies primarily relied on supervised learning strategies that utilize synthesized metal artifacts for training neural networks. However, the inherent discrepancies in synthesized data often hinder these methods' applicability to real-world, clinical settings. This paper addresses the gap by proposing an unsupervised learning framework, capitalizing on the concept of artifact disentangling within a neural architecture, named the Artifact Disentanglement Network (ADN).

Methodology

ADN is designed to independently encode the anatomical content and metal artifacts from CT images into separate latent spaces. The content is encoded from both artifact-affected and artifact-free images, allowing for the synthetic generation or removal of artifacts in images through decoding from these latent encodings. This design enables several image operations: transforming an artifact-affected image into an artifact-free version, the reverse image translation, and maintaining structural integrity in image reconstructions. Through a combination of adversarial learning, and specialized losses—artifact consistency, reconstruction, and self-reduction—the model obviates the need for paired training data, which is typically difficult to obtain in clinical practice.

Performance Evaluation

The authors tested ADN across a synthesized dataset (SYN) and two clinical datasets (CL1 and CL2). In the synthesized dataset, ADN demonstrated results comparable to the state-of-the-art supervised methods, measured in PSNR and SSIM. The qualitative assessment on the clinical datasets revealed ADN's superiority in generalizing to real-world scenarios where supervised methods, trained on synthesized data, struggled. Notably, in clinical contexts where projection data aren't available, ADN's CT-image-based post-processing approach shows particular prowess by omitting reliance on raw projection data.

Implications and Future Work

ADN's unsupervised methodology presents several practical implications for the medical imaging field. It can be directly applied to metal artifact reduction without requiring proprietary access to projection data or CT reconstruction algorithms. Moreover, since ADN's framework is adaptable, it can be extended to other artifact reduction challenges such as deblurring and denoising. By examining this unsupervised approach, the theoretical landscape of metal artifact reduction can evolve to further explore latent encoding techniques and their application in other complex noise reduction tasks.

Unsupervised learning frameworks like ADN indicate a methodological shift in medical image processing, showing the potential to simplify data preparation requirements while improving cross-domain applicability. Future research could focus on optimizing ADN's architecture and loss functions to enhance its robustness, as well as scaling its application to higher dimensional data or integrating it within broader diagnostic systems. Through such developments, ADN could spur advancements in both the practical deployment and theoretical modeling of artifact reduction in medical imaging.