Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Quad-Net: Quad-domain Network for CT Metal Artifact Reduction

Zilong Li, Qi Gao, Yaping Wu, Chuang Niu, Junping Zhang, Meiyun Wang, Ge Wang, Hongming Shan

TL;DR

Quad-Net addresses CT metal artifact reduction by extending dual-domain MAR to a quad-domain framework that jointly exploits sinogram, image, and their Fourier-domain information. The architecture combines SFR-Net for sinogram-Fourier restoration, LU-Net for local image refinement, and IFR-Net for cross-domain image-Fourier refinement, enabled by fast Fourier convolution to provide global context with minimal overhead. The approach achieves superior quantitative and visual performance over state-of-the-art methods, exhibits robustness to imprecise metal masks, and demonstrates practical value on clinical data, with code publicly available. This work highlights the importance of global receptive fields and Fourier-domain processing in MAR and paves the way for efficient, high-fidelity artifact mitigation in CT imaging.

Abstract

Metal implants and other high-density objects in patients introduce severe streaking artifacts in CT images, compromising image quality and diagnostic performance. Although various methods were developed for CT metal artifact reduction over the past decades, including the latest dual-domain deep networks, remaining metal artifacts are still clinically challenging in many cases. Here we extend the state-of-the-art dual-domain deep network approach into a quad-domain counterpart so that all the features in the sinogram, image, and their corresponding Fourier domains are synergized to eliminate metal artifacts optimally without compromising structural subtleties. Our proposed quad-domain network for MAR, referred to as Quad-Net, takes little additional computational cost since the Fourier transform is highly efficient, and works across the four receptive fields to learn both global and local features as well as their relations. Specifically, we first design a Sinogram-Fourier Restoration Network (SFR-Net) in the sinogram domain and its Fourier space to faithfully inpaint metal-corrupted traces. Then, we couple SFR-Net with an Image-Fourier Refinement Network (IFR-Net) which takes both an image and its Fourier spectrum to improve a CT image reconstructed from the SFR-Net output using cross-domain contextual information. Quad-Net is trained on clinical datasets to minimize a composite loss function. Quad-Net does not require precise metal masks, which is of great importance in clinical practice. Our experimental results demonstrate the superiority of Quad-Net over the state-of-the-art MAR methods quantitatively, visually, and statistically. The Quad-Net code is publicly available at https://github.com/longzilicart/Quad-Net.

Quad-Net: Quad-domain Network for CT Metal Artifact Reduction

TL;DR

Quad-Net addresses CT metal artifact reduction by extending dual-domain MAR to a quad-domain framework that jointly exploits sinogram, image, and their Fourier-domain information. The architecture combines SFR-Net for sinogram-Fourier restoration, LU-Net for local image refinement, and IFR-Net for cross-domain image-Fourier refinement, enabled by fast Fourier convolution to provide global context with minimal overhead. The approach achieves superior quantitative and visual performance over state-of-the-art methods, exhibits robustness to imprecise metal masks, and demonstrates practical value on clinical data, with code publicly available. This work highlights the importance of global receptive fields and Fourier-domain processing in MAR and paves the way for efficient, high-fidelity artifact mitigation in CT imaging.

Abstract

Metal implants and other high-density objects in patients introduce severe streaking artifacts in CT images, compromising image quality and diagnostic performance. Although various methods were developed for CT metal artifact reduction over the past decades, including the latest dual-domain deep networks, remaining metal artifacts are still clinically challenging in many cases. Here we extend the state-of-the-art dual-domain deep network approach into a quad-domain counterpart so that all the features in the sinogram, image, and their corresponding Fourier domains are synergized to eliminate metal artifacts optimally without compromising structural subtleties. Our proposed quad-domain network for MAR, referred to as Quad-Net, takes little additional computational cost since the Fourier transform is highly efficient, and works across the four receptive fields to learn both global and local features as well as their relations. Specifically, we first design a Sinogram-Fourier Restoration Network (SFR-Net) in the sinogram domain and its Fourier space to faithfully inpaint metal-corrupted traces. Then, we couple SFR-Net with an Image-Fourier Refinement Network (IFR-Net) which takes both an image and its Fourier spectrum to improve a CT image reconstructed from the SFR-Net output using cross-domain contextual information. Quad-Net is trained on clinical datasets to minimize a composite loss function. Quad-Net does not require precise metal masks, which is of great importance in clinical practice. Our experimental results demonstrate the superiority of Quad-Net over the state-of-the-art MAR methods quantitatively, visually, and statistically. The Quad-Net code is publicly available at https://github.com/longzilicart/Quad-Net.
Paper Structure (30 sections, 13 equations, 10 figures, 6 tables, 1 algorithm)

This paper contains 30 sections, 13 equations, 10 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional Methods for MAR
  4. Deep Learning-based MAR
  5. Fourier Neural Network
  6. Method
  7. Problem Formulation
  8. Overview of the Proposed Quad-Net
  9. Sinogram-Fourier Restoration Network (SFR-Net)
  10. Image-Fourier Refinement Network (IFR-Net)
  11. Total Objective Function
  12. Experimental Results
  13. Experimental Setup
  14. Datasets
  15. Evaluation metrics
  16. ...and 15 more sections

Figures (10)

  • Figure 1: The illustration of (a) local MAR and (b) global MAR in the sinogram. When employing conventional convolution for MAR in (a), the resulting receptive field is local and limited, which can only utilize neighboring angles within the receptive field to restore the corrupted signal (red star) corresponding to the interested pixel (blue point). In contrast, Fourier domain can easily provide global receptive field to predict the corrupted signal via all the angles that contain sufficient information about the interested pixel. Our idea is to synergize them for better sinogram restoration.
  • Figure 2: The framework of Quad-Net for MAR. (a) Our Quad-Net is to synergize all the features in the sinogram, image, and their corresponding Fourier domains, which takes little additional computational cost, and works across the four receptive fields to learn both global and local features as well as their relations. (b) The detailed network of SFR-Net in the sinogram domain and its Fourier domain. (c) Conventional LU-Net for stable base image refinement. (d) IFR-Net takes both an image and its Fourier spectrum to improve images using cross-domain contextual information.
  • Figure 3: Three different types of MAR in the sinogram domain: (a): sinogram completion with binary metal trace; (b): sinogram enhancement with binary metal trace; and (c): sinogram enhancement with metal mask projection. Both binary metal trace and metal mask projection offer an attentive mask for sinogram restoration.
  • Figure 4: Visual comparison of sinogram restoration results in the Fourier domain: (a) the reference metal trace; (b) the reference metal-corrupted sinogram; and the Fourier amplitude of (c) metal-corrupted sinogram; (d) LI; (e) DuDoNet; (f) DuDoNet++; (g) Quad-Net; and (h) Ground Truth.
  • Figure 5: Visual results of different configurations of Quad-Net: (a) Metal-corrupted images (inputs); (b) SFR-Net+LI+IFR-Net; (c) SR-Net+LU-Net+IR-Net; (d) Quad-Net; and (e) Ground Truth. The display window is (WL: 50 HU, WW: 500 HU).
  • ...and 5 more figures