PaperlayerView exact revision on arXiv

Source-linked AI summary

Convolutional Neural Network Based Metal Artifact Reduction in X-ray Computed Tomography

Yanbo Zhang, Hengyong Yu

arXiv:1709.01581v2physics.med-ph

TL;DR

Metal implants produce severe CT artifacts, while existing MAR methods can introduce distortions or inaccurate priors. This paper proposes CNN-MAR, which fuses uncorrected and pre-corrected images, constructs a tissue-processed CNN prior, and replaces affected projections before FBP; simulations and clinical applications show reduced artifacts and restored nearby structures.

  • Problem

    Metal artifacts remain a challenging x-ray CT problem because implants produce severe artifacts and existing methods can distort nearby structures or generate inaccurate prior images.

  • Method

    CNN-MAR trains on metal-free, metal-inserted, and corrected images, fuses uncorrected, BHC, and LI inputs, then uses a tissue-processed CNN prior for projection replacement and FBP.

  • Results

    Numerical simulations and clinical applications demonstrate significant artifact reduction and restoration of fine structures near metals.

  • Takeaways & Limitations

    The framework preserves tissue features while suppressing artifacts across simulated and clinical metal-implant cases.

Abstract

from arXiv · show

In the presence of metal implants, metal artifacts are introduced to x-ray CT images. Although a large number of metal artifact reduction (MAR) methods have been proposed in the past decades, MAR is still one of the major problems in clinical x-ray CT. In this work, we develop a convolutional neural network (CNN) based open MAR framework, which fuses the information from the original and corrected images to suppress artifacts. The proposed approach consists two phases. In the CNN training phase, we build a database consisting of metal-free, metal-inserted and pre-corrected CT images, and image patches are extracted and used for CNN training. In the MAR phase, the uncorrected and pre-corrected images are used as the input of the trained CNN to generate a CNN image with reduced artifacts. To further reduce the remaining artifacts, water equivalent tissues in a CNN image are set to a uniform value to yield a CNN prior, whose forward projections are used to replace the metal-affected projections, followed by the FBP reconstruction. The effectiveness of the proposed method is validated on both simulated and real data. Experimental results demonstrate the superior MAR capability of the proposed method to its competitors in terms of artifact suppression and preservation of anatomical structures in the vicinity of metal implants.

I. INTRODUCTION

Metal implants create severe CT artifacts, and existing correction methods remain challenging because artifact effects can corrupt tissue priors and distort nearby structures. The paper therefore develops a CNN-based framework that fuses uncorrected and pre-corrected images, trained using simulated metal-artifact databases.

  • Metal implants cause beam hardening, photon starvation, scatter, and strong star-shaped or streak artifacts in reconstructed CT images.
  • Existing MAR methods include physical-effects correction, projection-domain interpolation, prior-image methods, and iterative reconstruction, but each faces limitations.
  • LI can introduce artifacts and distort structures near large metals, while inaccurate tissue classification can produce false structures in NMAR prior images.
  • The proposed CNN framework trains on metal-free, metal-inserted, and corrected CT images, then fuses uncorrected, BHC, and LI images to generate a corrected CNN image.The database is created from clinical CT images with simulated metal artifacts and correction outputs.
  • The method further creates a CNN prior by processing tissue regions and uses its forward projections to replace metal-affected projections before FBP reconstruction.

2) Simple Metal Artifact Reduction:

The study uses LI and BHC as simple, fast pre-correction methods whose different correction mechanisms provide complementary inputs for the CNN.

  • 2) Simple Metal Artifact Reduction:: LI identifies metal-affected projections and replaces them with interpolated neighboring projections.
  • 2) Simple Metal Artifact Reduction:: LI and BHC are fast, easy to implement, require no manually selected parameters, and suppress artifacts through different schemes.
  • 2) Simple Metal Artifact Reduction:: BHC estimates metal contributions using the relation between metal path length and projection differences, then subtracts a fitted correction curve.
  • 2) Simple Metal Artifact Reduction:: BHC and LI corrected images are included as database samples, with metals not inserted back into the LI images.

B. Training a Convolutional Neural Network (CNN)

The CNN is trained to map three-channel patches containing uncorrected, BHC-corrected, and LI-corrected images to corresponding metal-free image patches.

  • B. Training a Convolutional Neural Network (CNN): Each database sample combines uncorrected, BHC, and LI images into a three-channel CNN input.
  • B. Training a Convolutional Neural Network (CNN): Input patches of size s × t × 3 and target patches of size s × t are extracted from corresponding image locations.
  • B. Training a Convolutional Neural Network (CNN): The database samples are randomly divided into training and validation groups before patch extraction.
  • B. Training a Convolutional Neural Network (CNN): The CNN learns a function H: R^(s×t×3) → R^(s×t) by minimizing a Frobenius-norm cost over input-target patch pairs.
  • B. Training a Convolutional Neural Network (CNN): The network uses five convolutional layers, with ReLU activations after the first four layers.

III. CNN-MAR METHOD

CNN-MAR combines metal segmentation, LI and BHC pre-correction, CNN-based image generation, tissue-prior construction, and projection replacement followed by FBP.

  • III. CNN-MAR METHOD: CNN-MAR first segments metal traces and applies LI and BHC to reduce artifacts before CNN processing.
  • III. CNN-MAR METHOD: The trained CNN generates an artifact-reduced image from the uncorrected and pre-corrected images.
  • III. CNN-MAR METHOD: Tissue processing converts the CNN image into a CNN prior, whose forward projection replaces metal-affected projection data.
  • III. CNN-MAR METHOD: The final reconstruction uses filtered backprojection after replacing the affected projections with the CNN prior's forward projection.

A. CNN Processing

The CNN combines uncorrected and pre-corrected CT images as a three-channel input to produce an artifact-reduced CNN image. The processed image suppresses artifacts that remain or are introduced by individual corrections.

  • The original uncorrected, BHC-corrected, and LI-corrected images are combined into a three-channel CNN input.
  • The CNN markedly suppresses artifacts in the processed image, including artifacts that persist after or are introduced by LI correction.

B. Tissue Processing

Tissue processing converts the CNN image into a CNN prior by flattening water-equivalent regions while preserving bone and smoothing tissue boundaries. Forward projection of this prior provides a substantially better estimate of missing projections than LI alone.

  • Tissue Processing: Water-equivalent tissues are assigned a uniform value to remove most remaining artifacts and form the CNN prior.
  • Tissue Processing: The CNN image is segmented into bone, water, and air using k-means clustering with safeguards for sparse or low-attenuation bone.
  • Tissue Processing: A five-pixel transition between water and other tissues reduces boundary discontinuities and artificial edges caused by constant-value replacement.
  • Tissue Processing: Metal pixels are replaced with nearest pixel values to avoid potential discontinuities at metal boundaries.
  • Tissue Processing: The resulting prior has flat water-equivalent regions while preserving bony structures, supporting projection interpolation.
  • Tissue Processing: Forward projection of the CNN prior produces a surrogate sinogram extremely close to the ideal sinogram, unlike LI interpolation.

IV. EXPERIMENTS

The experiments use simulated CT cases covering varied implant geometries and materials, with CNN training based on targeted artifact-containing patches. The CNN architecture preserves input size through padded 3 × 3 convolutions.

  • IV. EXPERIMENTS: The database contains 100 simulated cases built from 74 metal-free CT images and 15 metal shapes representing diverse clinical implants.
  • IV. EXPERIMENTS: The experiments include comparisons of CNN images and CNN priors, as well as sinogram-completion visualizations with enlarged ROI views.
  • IV. EXPERIMENTS: Simulations vary implant sizes, angles, positions, and materials, including titanium, iron, copper, and gold, to approximate clinical cases.
  • IV. EXPERIMENTS: The simulated acquisition assumes an equi-angular fan-beam geometry and a 120 kVp x-ray source.
  • IV. EXPERIMENTS: Each CNN layer uses a 3 × 3 convolutional kernel with padding 1, keeping the output image size equal to the input size.
  • IV. EXPERIMENTS: Training uses 10,000 64 × 64 patches, emphasizing patches with the strongest artifacts and supplementing them with random samples.

C. Numerical Simulation

Across three simulated metal-artifact cases, CNN-MAR generally provided the strongest combination of artifact suppression and anatomical-structure preservation, with the best reported RMSE and SSIM performance.

  • The three held-out cases included bilateral hip prostheses, fixation screws with an inserted shoulder-blade metal, and multiple dental fillings.
  • In the hip-prosthesis case, NMAR1 propagated false prior structures, LI blurred nearby bone, and CNN-MAR was evaluated alongside these alternatives.The corresponding prior images for NMAR1, NMAR2, and CNN-MAR were shown separately.
  • For the shoulder-blade case, CNN restored most nearby bony features without new artifacts, and CNN-MAR was very close to the reference.
  • For dental fillings, CNN preserved tissue features while avoiding most strong artifacts, and CNN-MAR achieved the best image quality.
  • CNN-MAR produced the smallest RMSEs for all three cases, while CNN images had accuracy comparable to NMAR2.RMSE was computed against reference images with metallic pixels excluded.
  • CNN-MAR achieved the highest SSIM in all three cases, indicating the strongest reported structural similarity among the compared methods.The paper describes higher SSIM as indicating better image quality.

B. Clinical Application

In a clinical head-CT case with a surgical clip, CNN-MAR most effectively reduced artifacts while preserving diagnostically relevant structures, whereas omitting tissue processing left tangent streaks.

  • The clinical evaluation used a head CT image containing a surgical clip, with corrected results compared across BHC, LI, NMAR1, NMAR2, and CNN-MAR.
  • The original, BHC, and LI images contained artifacts too strong to provide bleeding information, while CNN-MAR achieved the best image quality.
  • CNN-MAR left only one tiny dark streak, allowing the bright hemorrhage to be observed clearly and indicating potential for post-clipping diagnostic tasks.
  • Using CNN images directly as priors alleviated some artifacts but preserved most streaks tangent to the metals; tissue processing retained major structures and reduced this residual pattern.
  • In the presence of large metals, low-contrast features near metal may still be missing or distorted.

2) Selection of Input Images (MAR Methods):

The study evaluates how CNN-MAR performance changes with input-image channels, CNN architecture, and training-data size. More training data improves performance strongly, while larger architectures offer only slight gains at substantially higher computational cost.

  • Selection of Input Images (MAR Methods): The default three-channel input combines original, BHC, and LI images, while alternatives use five, two, or one channel.The five-channel input adds NMAR1 and NMAR2; the two-channel input uses original and LI, and the one-channel input uses NMAR2.
  • Selection of Input Images (MAR Methods): NMAR1 and NMAR2 provide limited new information because both derive from priors based on the original and LI images.
  • Architecture of the CNN: Larger CNNs yield slightly smaller RMSE and greater SSIM as convolutional layers, filters per layer, or filter size increase.The default CNN uses 5 convolutional layers, 32 filters per layer, and 3 × 3 filters.
  • Architecture of the CNN: The study employs a medium-size CNN because increasing architecture parameters considerably raises computational cost despite modest quality gains.
  • Training Data: Increasing training patches from 100 to 500, 2000, and 10000 dramatically decreases RMSE and increases SSIM, showing strong dependence on training-data size.

5) Training Epochs:

Training longer steadily improves CNN image quality, while tissue processing further improves CNN-MAR images. The framework’s effectiveness depends on suitable MAR inputs and varied training data, with segmentation accuracy and 3D training cost defining practical boundaries.

  • Training Epochs: RMSE decreases steadily and SSIM increases constantly as training epochs rise from 100 to 200, 1000, and 2000.After tissue processing, CNN-MAR image quality is further improved.
  • Training Epochs: CNN and tissue processing are mutually beneficial: CNN fuses information to avoid strong artifacts, while tissue processing removes moderate residual artifacts.
  • Training Epochs: Appropriate MAR methods provide information for distinguishing tissue structures from artifacts, while varied training cases improve the generality of the trained CNN.
  • Limitations: Inaccurate metal segmentation can compromise artifact-reduction performance, although advanced segmentation schemes can be applied to the method.
  • Limitations: Extending CNN-MAR from 2D slices to 3D volumes may improve performance but requires more training time.
  • Conclusion: Numerical simulations and clinical applications show significant artifact reduction and substantial restoration of fine structures near metals.
Loading 1709.01581v2…