Click here to


Are you sure ?

Yes, do it No, cancel

Real-Time Delineation of Implanted Cardioverter-Defibrillators and Metal Artifact Reduction in Cardiac Cone-Beam CT

I Park1,3*, J Chun1,3, S Olberg1,2, B Cai1, G Hugo1, C G Robinson1, J Kim3, S Mutic1, J Park1,2,3, (1) Department of Radiation Oncology, Washington University in St. Louis, St Louis, MO, (2) Department of Biomedical Engineering, Washington University in St. Louis, St Louis, MO, (3) Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, South Korea


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

Room: AAPM ePoster Library

Purpose: Implanted cardioverter-defibrillators (ICDs) in patients undergoing image-guided radiotherapy (IGRT) lead to severe streaking artifacts in cone-beam CT (CBCT) due to metallic components that degrade the overall image quality as well as treatment setup accuracy. Metal artifact reduction (MAR) methods address this issue by tracking metal-affected projections and replacing the metallic region through interpolation or inpainting. The rapid, unpredictable, and aperiodic motion of the heart, however, renders the tracking of ICDs in MAR methods impractical. In this study, we present a deep learning (DL)-based approach to track and segment ICDs in real-time and reconstruct MAR CBCT based on a GPU-accelerated projection inpainting and reconstruction framework.

Methods: Real-time ICD segmentation for each cone-beam projection was achieved using the fully convolutional DenseNet (FC-DenseNet) trained from 12 previously treated patients’ data containing ICDs, each with 894 cone-beam projections. Prior to training, ICDs were manually labeled at each projection angle with a generic feature extraction algorithm and manual refinement. Next, the MAR CBCT was reconstructed by inpainting the segmented regions of ICDs and using a Feldkamp-Davis-Kress (FDK)-based algorithm. To maximize the processing speed and efficiency, the code was implemented with GPU support for parallel processing purposes.

Results: The results indicate that FC-DenseNet can precisely delineate ICDs with minimal errors. The ICD segmentation accuracy was measured with a Dice similarity coefficient (DSC) of 0.72 and the processing time for each projection was 0.075 seconds. Using MAR, the metallic streaking artifacts of ICDs were significantly reduced in the volumetric CBCT without compromising image contrast and surrounding soft tissue visibility.

Conclusion: The proposed DL-based segmentation of ICDS in CBCT projections for MAR overcomes conventional limitations to effectively delineate ICDs subject to rapid, unpredictable, and aperiodic cardiac motion in order to remove severe streaking artifacts observed in CBCT during IGRT.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email