Collection of deep learning-based metal artifact reduction (MAR) articles for CT/CBCT Imaging.
Implicit neural representation-based method formetal-induced beam hardening artifact reduction in X-rayCT imaging 
H. S. Park et al.
Summary
This paper introduces a parameter-free metal-induced beam hardening correction (MBHC) method for CT imaging utilizing implicit neural representations (INRs). In contrast to the MBHC approach, the proposed method generates two tomographic images: one representing the monochromatic attenuation distribution and the other capturing nonlinear beam hardening effects. This eliminates the need for explicit metal segmentation and parameter estimation. The INR-generated images are combined to simulate forward projection data, which is then used to compute a loss function for network training. Experimental results demonstrate effective reduction of beam hardening artifacts arising from interactions between metals, bone, and teeth. Furthermore, the method exhibits potential for addressing challenges associated with photon starvation and truncated fields-of-view, leading to improved image reconstruction quality.
Medical Physics, 2025. [doi]
Metal implant segmentation in CT images based on diffusion model 
K. Xie et al.
Summary
The authors propose to use a diffusion model "DiffSeg" to segment the metal implants in CT images. The method is compared against U-Net, Attention U-Net, R2U-Net, and DeepLabV3+. The training and testing is conducted using simulated data. Additionally, the method is also tested on real clinical data. The method is shown to outperform the other methods. Better segmentation of metals is shown to reduce the artifacts when using NMAR.
BMC Medical Imaging, 2024. [doi]
A denoising diffusion probabilistic model for metal artifact reduction in CT
G. M. Karageorgos et al.
Summary
The authors propose to use a denoising diffusion probabilistic model (DDPM) for metal artifact reduction in CT. The model was used to inpaint the metal traces in the projection domain. The training was done in the unsupervised manner but during the inference, the segmented metal trace was required. The DDPM model was compared with a Partial Convolutions-based U-Net and a Gated Convolutions-based GAN. The proposed method showed promising results, however, the method was computationally expensive. Moreover, it was less effective in reducing met artifacts caused by large metal objects. The authors also noted that it is crucial to have an accurate metal trace segmentation for the best performance of the proposed network
IEEE TMI, 2024. [doi]
PND-Net: Physics-inspired Non-local Dual-domain Network for Metal Artifact Reduction
J. Xia et al.
Summary
The authors propose to use three networks. One is called non-local sinogram decomposition network (NSD-Net) to acquire the weighted artifact component in sinogram domain. The second network is an image restoration network (IR-Net) to reduce the residual and secondary artifacts in the image domain. The third trainable fusion network (F-Net) in the artifact synthesis path is used for unpaired learning. The method is compared with linear interpolation, NMAR, DuDoNet, DuDoNet++, InDuDoNet+, ADN, and U-DuDoNet.
IEEE TMI, 2024. [doi]
Dual Domain Diffusion Guidance for 3D CBCT Metal Artifact Reduction
Y. Choi et al.
Summary
The authors propose to use two 2D diffusion models in image domain to synthesize metal artifacts and metal free images, respectively. The combination of metal artifact and metal free images is forward projected and a guidance from the error in the projection domain is used. The results are compared with InDuDoNet+, ACDNet, FEL, and Blind DPS methods.
WACV, 2024. [doi]
Deep learning based projection domain metal segmentation for metal artifact reduction in cone beam computed tomography
H. Agrawal et al.
Summary
The authors proposed to use noisy Monte Carlo simulations to train a U-Net for metal segmentation in the projection domain. Including both the crops and full size images in the training data improved the metal segmentation performance. The evaluations were conducted on real clinical CBCT data to show the reduction in metal artifacts after inpainting the segmented metal traces in the projections. The evaluations included challenging datasets with large high density objects, motion artifacts, and out of field of view metals.
IEEE Access, 2023. [doi]
Quad-Net: Quad-domain Network for CT Metal Artifact Reduction
Z. Li et al.
Summary
In addition to the dual domain artifact reduction in sinogram and image domain, the paper proposed to also apply Fourier domain processing of features in sinogram and image domain, respectively.
[arXiv]. IEEE TMI, 2024. [doi].
Unsupervised metal artifacts reduction network for CT images based on
efficient transformer
L. Zhu et al.
Summary
A method to incorporate efficient Transformer blocks in the two generators of the CycleGAN. Additionally, the method also includes a loss term for the forward projections in the non-metal area. The results are compared with linear interpolation, CycleGAN and ADN.
BSPC, 2023. [doi]
Simulation-driven training of vision transformers enables
metal artifact reduction of highly truncated CBCT scans
F. Fan et al.
Summary
The paper proposes to use Swin-transformer for metal segmentation in the projection domain for CBCT data. Simulation data is used for the training. U-Net is shown to work best when the test data is similiar to the train data. But when the test data source is different than training, Swin transformer performs better.
Medical Physics, 2023. [doi]
DL-based inpainting for metal artifact reduction for cone beam CT using metal path length information
T. M. Gottschalk et al.
Summary
The authors propose to use a U-Net for inpainting the metal traces in the projection domain. The metal path length information is used to guide the inpainting along with the metal mask. Authors suggest that it is benificial to not dicard the information inside the metal mask. Additionally, 10 nearby projections were also concatenated to the input, making the total number of channels in the input 13. The output was the projection to be inpainted/corrected.
Medical Physics, 2022. [doi]
A fidelity-embedded learning for metal artifact reduction in dental CBCT 
H. S. Park et al.
Summary
The authors propose using an iterative method for fidelity-embedded learning to reduce metal artifacts in dental CBCT images. In each iteration, the network learns to predict the remaining metal artifact within the image. Additionally, a data-fidelity error term is calculated in the projection domain, but only for the non-metal areas. The Astra Toolbox was utilized for implementation. Training was conducted on simulated data, whereas testing was performed on both simulated and real clinical data.
Medical Phyics, 2022. [doi]
Metal Artifact Reduction In Cone-Beam Extremity Images Using Gated Convolutions
H. Agrawal et al.
Summary
The authors propose to use Gated Convoltuions for the metal area inpainting in the projection domain. The method is compared against linear interpolation, U-Net, and Partial Convolutions.
ISBI, 2021. [doi]
ADN: Artifact Disentanglement Network for
Unsupervised Metal Artifact Reduction
H. Liao et al.
Summary
The authors propose to use a disentanglement network to separate the metal artifact and the non-metal artifact latents. Two encoders are used to conde the latents. The latents are used by the decoders to generate the metal artifact affected and metal artifact free images. The method is compared against linear interpolation, NMAR, U-Net, cGANMAR,and CycleGAN among other methods. U-Net had the best PSNR and cGANMAR had the best SSIM. ADN had best PSNR and SSIM among the compared unsupervised methods.
[arxiv]. IEEE TMI, 2019. [doi]