Reducing CT Metal Artifacts by Learning Latent Space Alignment with Gemstone Spectral Imaging Data

TL;DR

Addresses CT metal artifacts by aligning ordinary CT slices $S_n$ with gemstone spectral imaging slices $S_g$ in latent space. The Latent GSI Alignment Framework (LGA) uses a specialized Artifact-reducing VAE, an Alignment Network with implant-type and metal-mask priors, and an Information Invariant Loss to preserve structure while removing artifacts. A new ArtifactGSI dataset of real patient data is released to train and benchmark methods across semantic segmentation, image quality, and expert evaluation. Results show substantial artifact reduction, high structural fidelity near or surpassing GSI, and improved clinical readability over prior approaches.

Abstract

Metal artifacts in CT slices have long posed challenges in medical diagnostics. These artifacts degrade image quality, resulting in suboptimal visualization and complicating the accurate interpretation of tissues adjacent to metal implants. To address these issues, we introduce the Latent Gemstone Spectral Imaging (GSI) Alignment Framework, which effectively reduces metal artifacts while avoiding the introduction of noise information. Our work is based on a key finding that even artifact-affected ordinary CT sequences contain sufficient information to discern detailed structures. The challenge lies in the inability to clearly represent this information. To address this issue, we developed an Alignment Framework that adjusts the representation of ordinary CT images to match GSI CT sequences. GSI is an advanced imaging technique using multiple energy levels to mitigate artifacts caused by metal implants. By aligning the representation to GSI data, we can effectively suppress metal artifacts while clearly revealing detailed structure, without introducing extraneous information into CT sequences. To facilitate the application, we propose a new dataset, Artifacts-GSI, captured from real patients with metal implants, and establish a new benchmark based on this dataset. Experimental results show that our method significantly reduces metal artifacts and greatly enhances the readability of CT slices. All our code and data are available at: https://um-lab.github.io/GSI-MAR/

Figures (4)

• Figure 1: Comparison of Artifacts Reduction Pipelines. (a) Most previous methods rely on synthetic artifact data derived from clean CT sequences of patients without implants. Additionally, many methods use image generation algorithms, which may introduce extraneous information, potentially compromising the reliability of the resulting CT sequences. (b) In contrast, our method utilizes real artifact CT pairs for training, effectively bridging the domain gap. Our approach employs a representation alignment algorithm, maintaining information consistency. (c) We provide a comparison of inference results between our method and previous methods to illustrate the effectiveness of our approach.
• Figure 2: Illustration of the Proposed Latent Space Alignment Framework. The pipeline consists of four stages: Data Processing, VAE encoding, Latent Space Alignment, and VAE decoding.
• Figure 3: Illustration of the Proposed Artifact-reducing VAE and Alignment Network. (a) Artifact-reducing VAE: The VAE Encoder takes CT volumes as input, using data augmentation on the center slice to enhance information aggregation. Two VAE decoders are employed: one trained on ordinary data and another on clean data, decoupling the decoding representations for artifact-affected and clean images. (b) Alignment Network: The network employs an encoder-decoder structure, enhanced with additional input signals and transformer-residual modules.
• Figure 4: Qualitative Comparisons. (a) Comparison on the Test Set: Images of patients with hip prostheses used in total hip arthroplasty, fracture internal fixation, and spinal internal fixation. (b) Comparison on the Generalization Set: Evaluation on data from unseen CT machines (SpineWeb dataset, Siemens, Philips, and UIH CT machines) to demonstrate generalization.