Source-linked AI summary
DuDoNet: Dual Domain Network for CT Metal Artifact Reduction
Wei-An Lin, Haofu Liao, Cheng Peng, Xiaohang Sun, Jingdan Zhang, Jiebo Luo, Rama Chellappa, Shaohua Kevin Zhou
TL;DR
Metal implants create structured, non-local CT artifacts that image-domain methods struggle to model, while sinogram correction can introduce secondary artifacts through inconsistency. DuDoNet jointly refines sinograms and CT images through a trainable Radon inversion layer, and experiments report effective suppression of metal shadows with detail recovery and greater computational efficiency. Its formulation assumes a ground-truth artifact-free CT image for the cited training-related quantity.
Problem
Metal implants produce structured, non-local CT artifacts, while sinogram-domain methods can introduce severe secondary artifacts from sinogram inconsistency.
Method
DuDoNet jointly restores sinograms and CT images with SE-Net, a Radon inversion layer for gradient back-propagation, IE-Net, and a Radon consistency loss.
Results
DuDoNet effectively suppresses metal shadows, recovers CT details, and is computationally more efficient than the evaluated alternatives.
Takeaways & Limitations
Dual-domain enhancement provides a supported route to improving sinogram consistency while refining artifact-reduced CT images.
Takeaways & Limitations
The cited Radon inversion-layer formulation uses ground-truth CT images without metal artifacts.
Abstract
from arXiv · showhide
Computed tomography (CT) is an imaging modality widely used for medical diagnosis and treatment. CT images are often corrupted by undesirable artifacts when metallic implants are carried by patients, which creates the problem of metal artifact reduction (MAR). Existing methods for reducing the artifacts due to metallic implants are inadequate for two main reasons. First, metal artifacts are structured and non-local so that simple image domain enhancement approaches would not suffice. Second, the MAR approaches which attempt to reduce metal artifacts in the X-ray projection (sinogram) domain inevitably lead to severe secondary artifact due to sinogram inconsistency. To overcome these difficulties, we propose an end-to-end trainable Dual Domain Network (DuDoNet) to simultaneously restore sinogram consistency and enhance CT images. The linkage between the sigogram and image domains is a novel Radon inversion layer that allows the gradients to back-propagate from the image domain to the sinogram domain during training. Extensive experiments show that our method achieves significant improvements over other single domain MAR approaches. To the best of our knowledge, it is the first end-to-end dual-domain network for MAR.
1. Introduction
Metal artifacts are structured, non-local, and difficult to remove directly in images, while sinogram-domain correction can introduce secondary artifacts. DuDoNet addresses both issues through end-to-end dual-domain refinement linked by a Radon inversion layer.
- Metal artifacts are structured and non-local, making image-domain modeling extremely difficult.
- Sinogram-domain methods replace metal-corrupted regions, but physical inconsistency can create severe secondary artifacts.
- DuDoNet jointly learns sinogram restoration and CT-image enhancement using two CNNs in an end-to-end dual-domain network.
- The network combines SE-Net, a Radon inversion layer, and IE-Net; the inversion layer back-propagates image-domain gradients to the sinogram domain.
- Extensive experiments on CT images from hundreds of patients show superior artifact-reduced images from dual-domain enhancement.
2. Backgrounds and Related Works
CT reconstructs anatomical structure from X-ray projection data, but metallic implants add an ill-posed artifact term. Existing inpainting, iterative, and image-domain approaches face consistency, computational, or artifact-suppression limitations.
- CT reconstruction uses projection data to recover the anatomical attenuation map, but metallic implants introduce energy-dependent distortion and metal artifacts.
- Perfect MAR requires suppressing the artifact term without altering the reconstructed anatomy, an ill-posed problem because both affect the metal-trace region.
- Inpainting-based Methods: Inpainting methods treat metal-trace data as missing, but inconsistent replacements produce secondary artifacts after reconstruction.
- MAR by Iterative Reconstruction: Iterative reconstruction is computationally slow and handcrafted regularization can produce over-smoothed reconstructions.
- Image-domain CNN methods can reduce mild artifacts, but image-domain enhancement is insufficient for intense metal shadows.
3. Proposed Method
DuDoNet restores metal-corrupted sinograms and refines reconstructed CT images through a linked dual-domain pipeline. Its differentiable Radon inversion layer enables image-domain losses to back-propagate to sinogram enhancement while enforcing consistency.
- Sinogram Enhancement Network: The model takes a degraded sinogram and metal-trace mask, initializes masked values by linear interpolation, then restores them with SE-Net.SE-Net uses a mask pyramid U-Net to preserve small metal-trace information during down-sampling.
- Radon Inversion Layer: RIL reconstructs CT images from fan-beam sinograms using parallel-beam conversion, Ram-Lak filtering, and backprojection.The conversion and filtering modules provide a more memory-efficient implementation of fan-beam filtered backprojection.
- Radon Inversion Layer: Because the derivative of RIL with respect to its input is the projection operation, image-domain losses can be aggregated and projected into the sinogram domain.This property enables joint learning across sinogram and image domains.
- Image Enhancement and Objective: A Radon consistency loss penalizes reconstructed secondary artifacts, while IE-Net further refines the reconstructed CT image through residual learning.The consistency loss compares the reconstructed image against a ground-truth artifact-free CT image; IE-Net is optimized with an L1 loss.
4. Experimental Results
Experiments evaluate DuDoNet on simulated metal artifacts using component ablations and comparisons with existing MAR methods. The full dual-domain model consistently improves artifact suppression, detail recovery, quantitative quality, and runtime.
- Experimental Setup: The evaluation uses simulated artifacts on real patient CTs, incorporating polychromatic X-rays, partial volume effects, and Poisson noise.The dataset contains 4,000 training images from 320 patients and 200 test images from 12 patients, with 100 metal shapes.
- Ablation Study: At least 0.3 dB improvement across all metal sizes results from adding Radon consistency loss.The full model also better recovers organ shape visually than the configuration without RC loss.
- Ablation Study: Mask pyramid produces an approximately 0.2 dB gain for small metals, while performance is nearly identical for large metals.The architecture retains metal information across multiple scales and recovers fine details near small implants.
- Dual Domain Learning: Single-domain IE-Net and IE-Net-RDN fail to recover corrupted organ boundaries, whereas dual-domain refinement recovers boundaries and fine details.SE-Net reduces secondary artifacts in the reconstructed image before IE-Net performs image refinement.
- Input Sinogram: Using linearly interpolated sinograms and their reconstructed CT images suppresses artifacts more effectively than directly processing corrupted inputs.The alternative leaves metal shadows and streaking artifacts insufficiently suppressed.
- Comparison with State-of-the-Art Methods: DuDoNet achieves the best PSNR and SSIM consistently across all categories while combining artifact reduction with fine-detail retention.Visually, it reduces metal shadows while retaining details better than sinogram-inpainting and image-domain baselines across implant configurations.
- Runtime: DuDoNet runs almost 4× faster than RDN while achieving superior performance when linear-interpolation time is included.The Radon inversion layer takes 0.24 ms for reconstruction and 11.40 ms for gradient back-propagation on an Nvidia 1080Ti.
5. Conclusion
The paper concludes that DuDoNet jointly improves sinogram consistency and CT-image quality through dual-domain learning. Experiments show effective suppression of metal shadows, detail recovery, and greater computational efficiency than competing approaches.
- 5. Conclusion: DuDoNet jointly improves sinogram consistency and refines CT images using a Radon inversion layer, Radon consistency loss, and mask pyramid U-Net.The network is presented as an end-to-end trainable dual-domain refinement model for MAR.
- 5. Conclusion: Experiments demonstrate effective metal-shadow suppression and detail recovery despite limitations of sinogram-inpainting and very-deep image-domain networks.The conclusion also reports that the dual-domain model is computationally more efficient.
A. Fanbeam CT Geometry
Fanbeam CT geometry models a rotating X-ray source and arc detector around an object. Projection data are indexed by projection angle and detector position.
- A. Fanbeam CT Geometry: The X-ray source and arc detector rotate around the origin, with source-to-origin distance D.For each projection angle β, the detector records transmitted X-rays as a one-dimensional signal indexed by γ.
B.1. Network Architecture
DuDoNet uses sinogram and image enhancement networks, with a structured convolutional architecture for SE-Net and a U-Net-based IE-Net.
- SE-Net and IE-Net form the two main components of the proposed DuDoNet architecture.IE-Net uses an architecture identical to U-Net.
- The notation K#-C#-S#-P# specifies convolution kernel, channel, stride, and padding sizes.
- Sinogram masks are down-sized by factors of k through the Mt ↓k notation.
B.2. Radon Inversion Layer
The Radon inversion layer converts fan-beam sinograms, filters them, and backprojects the result into CT images while supporting gradient propagation.
- RIL contains parallel-beam conversion, Ram-Lak filtering, and backprojection modules.It converts Yfan(β, γ) into Ypara(t, θ), then reconstructs X(u, v).
- The conversion uses t and θ as detector location and projection angle in parallel-beam geometry.Two one-dimensional interpolations implement the variable changes for θ and t.
- The filtering module uses the discrete Fourier transform and inverse transform along the detector dimension.The implementation uses torch.fft and torch.ifft in PyTorch.
- The backprojection module takes filtered projections Q(t, θ) and reconstructs the CT image X(u, v).
- Backpropagation computes CT-image gradients with respect to the sinogram, and the backprojection module is implemented as a PyTorch CUDA extension.
C. Evaluation on CT Images with Real Metal Artifact
Evaluation on real metal-artifact CT images is difficult because clinical MAR processing obscures the baseline and sinogram data are generally unavailable. The authors construct evaluation inputs from DeepLesion images and report that DuDoNet reduces artifacts despite inconsistent imaging geometry.
- Clinical MAR algorithms and limited access to metal-corrupted sinograms make real-patient MAR evaluation challenging.
- The evaluation manually collects metal-artifact CT images from DeepLesion and derives metal traces and LI sinograms.
- Metal masks are obtained by thresholding the metal-corrupted CT image at 2,000 HU and forward-projecting with the manuscript’s imaging geometry.
- LI sinograms are synthesized from metal-corrupted sinograms and metal traces by linear interpolation.
- Inconsistent imaging geometry can make sinogram-based MAR methods produce worse visual quality than raw CT, whereas DuDoNet reduces artifacts in real CT images.
D. Additional Visual Comparisons on CT images with Synthesized Metal Artifact
The section discusses practical data requirements for deep-learning MAR and refers readers to additional visual comparisons on synthesized-metal-artifact CT images.
- Additional visual-comparison results are referenced on pages 13 and 14.
- CNNMAR and cGAN-CT require paired CT images with and without metallic implants from the same patient.
- The DuDoNet approach treats data within the metal trace as missing and replaces it using LI.