Click here to


Are you sure ?

Yes, do it No, cancel

Two-Stage Generative Adversarial Network (GAN) for Image-Based Metal Artifact Reduction

H Kim1,2*, C Kim2, B Cho3, J Kim4, (1) Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, KR, (2) Department of Radiation Oncology, Asan Medical Center, Seoul, KR


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

Room: AAPM ePoster Library

Purpose: Kilovoltage computed tomography (KVCT) images are susceptible to artifact from high density material, mostly metal. This work develops a metal artifact reduction (MAR) by utilizing distinct physical characteristics of imaging modalities and two-stage serial generative adversarial deep neural network (GAN).

Methods: Most of the conventional MAR techniques was achieved by modifying sinogram data: 1)finding metal-associated element in sinogram from binary masking, 2)vacating the elements and filling those with interpolation. The recent breakthrough from deep neural network focuses on how accurately estimates the blank elements instead of interpolating approach. Our proposed workflow promotes the image-based approach with an aid to deep neural network, specifically GAN, without proceesing sinogram. The major challenge in image-based MAR using deep neural network is to create pairs of KVCT images with and without artifacts for training. Referring to a fact that the other imaging modalities such as MVCT could be metal-artifact-resistant, the proposed work devised a two-stage serial deep neural network. The first network is to train MVCT (input) and KVCT (output), which generates the artifact-free KVCT by GAN. The second network conducts the metal artifact reduction by training the artifact-free KVCT derived from the first network against KVCT with metal artifacts.

Results: We trained the networks with 72 planning KVCT and MVCT scans, where the MVCTs were acquired by first fraction of treatment on TomoTherapy. 60 and 12 scans were used for training and testing the second network, respectively. Compared to KVCTs that passed through commercially available MAR software, our proposed method effectively reduced the metal artifacts for all testing cases. For more specific analysis, phantom experiments and effects on dose calculations would have to be investigated.

Conclusion: This study presented a novel technique for MAR in planning KVCT by using metal-artifact-resistant MVCT and two-stage deep neural network, which successfully eliminates the metal artifacts.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email