Click here to

Session: Deep Learning for Image Quality in Treatment Planning [Return to Session]

Deep Learning-Based Metal Artifact Reduction for Planning CT in Radiation Therapy for Head and Neck Cancer

J Lee1, J Lee2, S Jung3*, (1) Department of Nuclear Engineering, Ulsan National Institute of Science and Technology, Ulsan, Ulsan, KR, (2) Department of Computer Engineering, Ulsan National Institute of Science and Technology, KR, (3) Department of Radiation Oncology, Seoul National University Hospital, Seoul, KR


SU-H330-IePD-F5-6 (Sunday, 7/10/2022) 3:30 PM - 4:00 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 5

Purpose: To develop a deep learning-based metal artifact reduction (DL-MAR) technique for planning computed tomography (pCT) images in radiation therapy (RT) for head and neck cancer and to perform a dosimetric comparison between pCT images with DL-MAR and those with a commercial MAR (OMAR).

Methods: The whole dataset consisted of 120 kVp CT images of 35 patients who had no metallic implant and underwent RT for head and neck cancer from 2019 to 2020. The CT images for treatment planning were acquired by CT simulator (Brilliance Big Bore CT, Philips, Best, The Netherlands). Total 8,568 images were simulated based on the CT images through insertion of virtual metal shapes and manipulation of sinogram domain. To develop DL-MAR, a convolutional neural network with encoder-decoder structure was implemented and trained using the 6,426 simulated images. To evaluate the MAR performance of the trained model, pearson correlation coefficient (PCC), SSIM, and PSNR between the output and target images were calculated with remaining 2,142 simulated images. Dosimetric evaluation for one patient with metallic implants was conducted by using a treatment planning system (TPS).

Results: The trained CNN model showed superior metal artifact reduction performance for the simulated test images. Average PCC, SSIM, and PSNR scores from the simulated test images were 0.985, 0.998, and 41.998, respectively. Dark/white streak artifacts in output images from DL-MAR were well reduced. For dosimetric evaluation, no significant difference was observed between dose distributions on pCT images with DL-MAR and those with OMAR. Although the metal artifact was dramatically reduced and the image quality was improved, it was shown that pCT images with DL-MAR had no impact on calculated dose distributions.

Conclusion: The deep learning-based metal artifact reduction technique has been successfully developed and shown superior metal artifact reduction performance.


CT, Radiation Therapy, Image Artifacts


IM- CT: Metal artifact reduction

Contact Email