MENU

Click here to

Ă—

Are you sure ?

Yes, do it No, cancel

Self-Supervised Metal Artifact Reduction in X-Ray Computed Tomography by Joint Sinogram Completion and Image Refinement

L Yu*, Z Zhang, L Xing, Stanford Univ School of Medicine, Stanford, CA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: To alleviate the need for synthesized metal artifact pairs for network training, a generalizable deep-learning-based framework for CT metal artifact reduction has been proposed in a self-supervised learning fashion, combing the complementary benefit of sinogram domain completion and CT image domain refinement.


Methods: With a fan-beam geometry, we simulate the metal-free sinogram from the CT images and randomly simulate possible metal trace with the forward projection of a random metal mask. We first train a neural network to restore the metal trace region in the sinogram. We further proposed an FBP consistency loss to encourage the network to generate more geometry consistency results and adopt a residual learning-based image refinement module to remove most of the secondary-artifact. Thanks to the power of dual-domain deep learning, the CNN-out remove most of the secondary-artifacts. Finally, we produce a generated sinogram from the CNN refined image and incorporate the corresponding pixel in metal-affected regions into the origin metal sinogram, followed by the FBP to reconstruct the final metal-free image.


Results: Testing is conducted on 200 simulated images with additional metal masks. The root mean square error (RMSE) and structure similarity index (SSIM) of our method with the refer CT images are 34.34HU and 0.9770, which is better than normalized MAR (47.03HU and 0.9594). Visual results also indicate that our method can better alleviate artifacts while preserving the structure of the images.


Conclusion: This work indicates the potential of deep learning techniques for effective metal artifact reduction by combing sinogram domain and image domain processing simultaneously. Our experiments show that the generated images are qualitatively and quantitatively better than the traditional method. We also show the feasibility of adopting self-supervised learning techniques to reduce the number of synthesized training images. Future research includes evaluating our method on clinical CT images.

Keywords

CT, Image Correlation, Backprojection

Taxonomy

IM- CT: Metal artifact reduction

Contact Email