MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All
  4. Development of a CT Metal Artifact Reduction Technique Based On Deep Learning

Development of a CT Metal Artifact Reduction Technique Based On Deep Learning

J Lee1*, H Cho2 , H Kim3 , H Chae4 , S Ye1,5 , (1) Seoul National University, Seoul, Korea, (2) Interpark, Seoul, Korea, (3) Artificial Intelligence Research Institute, Seongnam-si, Korea, (4) Seoul National University Hospital, Seoul, Korea, (5) Seoul National University College of Medicine, Seoul, Korea

Presentations

(Tuesday, 7/31/2018) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 6

Purpose: To develop a novel MAR (Metal Artifact Reduction) technique based on Deep Learning to improve segmentation and dose calculation in CT

Methods: In order to develop Deep Learning based MAR, CT images of 300 patients with metal artifacts and O-MAR (Metal Artifact Reduction for Orthopedic Implants, Philips) images were obtained as pairs under IRB approval. We also acquired clear CT images of 300 patients without metal. MAG (Metal Artifact Generation) model was trained with O-MAR images as inputs and metal artifact images as targets by using FusionNet structure. Then, metal shapes extracted from metal artifact images were inserted into clear CT images, and artifact was generated in these metal inserted images through MAG model. Consequently, Metal Artifact Dataset which includes pairs of normal CT with metal insertion and the image with generated artifact was created. With this dataset, we could train MAR model to put the artifact generated images as inputs and the normal image with metal as targets. Finally, we compare the contour and dose calculation results from a metal artifact image and an artifact-reduced image using the MAR model.

Results: MAG model was trained with O-MAR and original metal artifact images and it was able to imitate the pattern of metal artifact. The optimized MAG model generated various image pairs for Metal Artifact Dataset. The dataset was utilized to train MAR model and optimization of the model is in progress.

Conclusion: A novel MAR technique based on Deep Learning is under development. The contour and dose calculation results from both metal artifact images and artifact-reduced images will be compared together to investigate performance of the MAR technique further.

Funding Support, Disclosures, and Conflict of Interest: This study was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT: Ministry of Science and ICT) (No. NRF-2017M2A2A6A01071214).

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Additional Links