PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Sinogram Completion with Image Prior for Metal Artifact Reduction in CT Images

Lequan Yu, Zhicheng Zhang, Xiaomeng Li, Lei Xing

arXiv:2009.07469v1eess.IVcs.CV

TL;DR

Metal implants cause severe CT artifacts that can impair diagnosis and radiation-therapy dose calculation. The paper addresses this by jointly training PriorNet and SinoNet for prior-guided residual sinogram completion, then reconstructing the result with FBP. Experiments on simulated and real artifacts report superior artifact reduction, anatomical-structure preservation, and performance over other MAR methods.

  • Problem

    Metal implants create missing projection data and severe CT artifacts, while existing MAR methods lack a standard clinical solution.

  • Method

    The framework jointly trains PriorNet for prior-image generation and SinoNet for residual sinogram completion, using differentiable reconstruction and FBP.

  • Results

    The method produces superior artifact-reduced results, preserves anatomical structures, and outperforms other MAR methods on simulated and real artifact data.

  • Takeaways & Limitations

    Prior-guided sinogram completion combines deep learning with conventional MAR processing while retaining geometry-consistent reconstruction through FBP.

  • Takeaways & Limitations

    The method was trained and evaluated on simulated projection data because public real projection data were unavailable.

Abstract

from arXiv · show

Computed tomography (CT) has been widely used for medical diagnosis, assessment, and therapy planning and guidance. In reality, CT images may be affected adversely in the presence of metallic objects, which could lead to severe metal artifacts and influence clinical diagnosis or dose calculation in radiation therapy. In this paper, we propose a generalizable framework for metal artifact reduction (MAR) by simultaneously leveraging the advantages of image domain and sinogram domain-based MAR techniques. We formulate our framework as a sinogram completion problem and train a neural network (SinoNet) to restore the metal-affected projections. To improve the continuity of the completed projections at the boundary of metal trace and thus alleviate new artifacts in the reconstructed CT images, we train another neural network (PriorNet) to generate a good prior image to guide sinogram learning, and further design a novel residual sinogram learning strategy to effectively utilize the prior image information for better sinogram completion. The two networks are jointly trained in an end-to-end fashion with a differentiable forward projection (FP) operation so that the prior image generation and deep sinogram completion procedures can benefit from each other. Finally, the artifact-reduced CT images are reconstructed using the filtered backward projection (FBP) from the completed sinogram. Extensive experiments on simulated and real artifacts data demonstrate that our method produces superior artifact-reduced results while preserving the anatomical structures and outperforms other MAR methods.

I. INTRODUCTION

Metal implants create missing projection data and severe artifacts that can impair diagnosis and radiation-therapy dose calculation. The paper addresses these limitations with jointly trained image- and sinogram-domain networks for sinogram completion.

  • Metal implants cause missing X-ray projection data and strong star-shaped or streak artifacts in reconstructed CT images.
  • Existing MAR methods lack a standard clinical solution and include sinogram interpolation, prior-image methods, and image-restoration networks.
  • Image-domain methods can introduce small anatomical structure changes because neural networks lack geometry constraints.
  • The proposed framework formulates MAR as deep sinogram completion using SinoNet, while PriorNet generates a prior image to guide completion.
  • Residual sinogram learning improves completion continuity and helps alleviate new artifacts in reconstructed CT images.
  • The framework jointly leverages image- and sinogram-domain MAR techniques and is evaluated on simulated and real metal artifacts.

II. METHODOLOGY

The framework combines linear-interpolation initialization, prior-image generation, forward projection, and sinogram completion in a joint image–sinogram pipeline. PriorNet refines corrupted and interpolated reconstructions, and its projected prior guides SinoNet.

  • The framework integrates image-domain prior-image generation with sinogram-domain completion for metal artifact reduction.
  • Linear interpolation first produces an initial corrected sinogram from the metal-affected sinogram and metal trace mask.
  • PriorNet refines the metal-corrupted and linearly interpolated reconstructions to generate a prior image with reduced artifacts.
  • PriorNet uses a reduced-channel U-Net and L1 loss against the artifact-free ground-truth CT image.
  • The prior image is forward-projected to obtain a prior sinogram that guides completion of missing projections.

C. Deep Sinogram Completion

SinoNet completes metal-affected projections by learning residual differences between the prior and interpolated sinograms. Mask-aware processing and auxiliary supervision support continuity and training efficiency.

  • SinoNet takes the interpolated sinogram, prior sinogram, and metal trace as inputs and restores projections within the metal trace.
  • Residual learning estimates correction values rather than absolute projections, reducing discontinuity at metal-trace boundaries.
  • The predicted output is composited with the interpolated sinogram using the metal trace mask to form the corrected sinogram.
  • A mask-pyramid U-Net preserves metal-trace information across network layers because the trace occupies a small sinogram region.
  • The training objective supervises both the corrected sinogram and the pre-composited output, with β empirically set to 0.1.

D. Overall Objective Function and Technical Details

The framework adds an FBP loss to address geometry consistency and new reconstructed-image artifacts, jointly trains PriorNet and SinoNet end-to-end, and uses simulated data for training.

  • The sinogram L1 loss alone does not enforce geometry consistency or penalize new artifacts in reconstructed CT images.
  • The FBP loss uses differentiable filtered backprojection to encourage geometry-consistent sinogram completion and reduce reconstructed-image artifacts.
  • PriorNet and SinoNet are jointly trained end-to-end with a total objective combining prior-image, sinogram, and FBP losses.
  • The loss weights α1 and α2 are empirically set to 1.0 in the experiments.
  • The framework uses original metal-affected sinograms and metal traces as inputs, with simulated metal masks providing training traces through forward projection.

A. Dataset and Simulation

The evaluation uses simulated artifacts generated from DeepLesion and additional head CT images, with fan-beam projections and specified image, sinogram, and training configurations.

  • The simulated dataset uses 1000 CT images and 90 metal masks for training, with 10 remaining masks paired with 200 additional CT images.
  • DeepLesion visual comparisons include reference images, simulated metal artifacts, and results from LI, NMAR, CNNMAR, cGANMAR, DuDoNet, and the proposed method.
  • Metal artifacts are simulated by inserting metallic implants into clean CT images while modeling beam hardening, Poisson noise, and partial volume effects.
  • The simulation uses fan-beam geometry with 640 projection views sampled uniformly over 0-360 degrees, producing 416×416 images and 641×640 sinograms.
  • Training uses end-to-end PriorNet and SinoNet optimization with differentiable FP and FBP operations, the Adam optimizer, 400 epochs, and batch size 8.

C. Experimental Results on DeepLesion Data

On DeepLesion data, the proposed method is compared with conventional interpolation and deep-learning MAR methods using quantitative metrics and visual results across simulated dental-filling cases.

  • The comparison includes LI, NMAR, CNNMAR, cGANMAR, and DuDoNet as conventional and deep-learning MAR baselines.
  • Visual experiments include head CT images with different numbers of simulated dental fillings and display the corresponding metal regions in red.
  • Compared with DuDoNet, the proposed method reduces RMSE by 6.85 HU and achieves slightly better SSIM values.
  • The proposed framework achieves the best DeepLesion performance in RMSE and SSIM among the compared methods.

2) Qualitative analysis:

Qualitative experiments show that the proposed method reduces artifacts while preserving structural detail across simulated cases, different implant sizes, head CT images, and real-metal examples.

  • Qualitative analysis: DeepLesion simulations contain severe streaking artifacts and dark strips between metal implants before correction.
  • Qualitative analysis: For small implants, the proposed method reduces most artifacts and retains fine structural details compared with competing methods.
  • Qualitative analysis: For large implants, competing interpolation and image-domain methods lose structural detail or introduce secondary artifacts, whereas the proposed method preserves more detail than CNNMAR.
  • Qualitative analysis: On head CT images with simulated dental fillings, the proposed method reduces artifacts and is comparable with CNNMAR trained on head CT images.
  • Qualitative analysis: Real-metal visual results show effective artifact reduction while preserving fine-grained anatomical structures.
  • Qualitative analysis: The proposed method remains reasonable under slightly over- or under-segmented metal masks, although over-segmentation requires completing more projection values.

E. Experiments on CT Images with Real Metal Artifacts

The framework is evaluated on clinical CT images with real metal artifacts using simulated projection data and varying metal masks. Accurate mask segmentation is important, although slight errors remain manageable with distinct residual effects.

  • Real metal artifact data: Clinical CT images with metal artifacts were evaluated by collecting scans, segmenting metal at 2000 HU, and forward-projecting the masks to generate metal traces.The resulting traces identify pixels affected by metal for evaluation.
  • Metal mask segmentation: Accurate metal trace masks are vital for good MAR performance, motivating comparisons across thresholding, dilation, and erosion-based masks.The study examines how segmentation errors influence the final artifact-reduced images.
  • Metal mask segmentation: Slight over-segmentation remains reasonable but completes more projection values and introduces additional shading artifacts.The network can handle enlarged traces without strong new artifacts, but image shading increases.
  • Metal mask segmentation: Under-segmentation leaves unreliable projection data inside the original metal trace, so only some artifacts are reduced and residual streaks remain.A narrower mask causes the completion network to reuse some corrupted measurements.

1) Effectiveness of prior image generation:

A generated prior image improves sinogram completion, and residual sinogram learning further improves artifact reduction on simulated and real examples. Tissue processing was less effective than the end-to-end strategy.

  • Effectiveness of prior image generation: Prior image generation improves sinogram completion, as direct deep completion without a prior produces higher RMSE and lower SSIM than the proposed method.This ablation verifies the contribution of the prior-image generation procedure.
  • Effectiveness of prior image generation: Generated prior images contain fewer artifacts than the original metal images, while final MAR results further reduce artifacts relative to those priors.The comparison is shown on DeepLesion and head CT data.
  • Effectiveness of prior image generation: Using only the original metal image as PriorNet input produces slightly worse simulated-dataset results than incorporating the LI-corrected image.The comparison indicates that LI-corrected input facilitates prior image generation.
  • Effectiveness of residual sinogram learning: Residual sinogram learning further reduces artifacts on simulated and real clinical samples and yields higher RMSE and lower SSIM than the residual-learning variant.The qualitative comparison appears in Fig. 9, with quantitative results reported in Table II.
  • Effectiveness of prior image generation: The tissue-processing variant remains inferior to end-to-end deep sinogram completion, which can learn to reduce mild prior-image artifacts.This comparison supports directly using the generated CNN output as the prior image.

IV. DISCUSSION

The framework combines image- and sinogram-domain learning through prior generation and residual completion, aiming to improve boundary continuity and generalization. Its main boundaries are simulated training data and dependence on accurate metal masks.

  • Discussion: The method formulates MAR as deep sinogram completion and combines a learned prior image with residual completion to improve projection continuity and reduce new artifacts.The framework is intended to retain deep-learning benefits while reducing overfitting risk.
  • Discussion: The final image is FBP-reconstructed from the completed sinogram, modifying only metal-trace values to preserve anatomy and avoid CNN image-resolution loss.This distinguishes the method from approaches that directly use image-domain refinement outputs.
  • Limitations: Training currently relies on simulated metal artifacts because paired metal-free and metal-inserted CT data are clinically impractical to acquire.The authors identify simulation quality as a factor that can influence network performance.
  • Limitations: The method was evaluated on simulated datasets, while real-data deployment still requires real projection evaluation and accurate metal segmentation for stable results.Future work includes improved metal identification or deep-learning-based segmentation.
  • Conclusion: The conclusion presents the framework as jointly trained prior generation and sinogram completion with potential for CT images containing real artifacts.Future directions include simultaneous metal-mask identification and unsupervised artifact reduction.
Loading 2009.07469v1…