Click here to


Are you sure ?

Yes, do it No, cancel

Convolutional Network Based Motion Artifact Reduction in Cone-Beam CT

P Paysan1*, A Strzelecki1, F Arrate2, P Munro1, S Scheib1, (1) Varian Medical Systems Imaging Laboratory, Daettwil AG,(2) Varian Medical Systems, Palo Alto, CA


(Tuesday, 7/16/2019) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 9

Purpose: Motion artifacts impair cone-beam CT (CBCT) image quality. A method to reduce the impact of streak artifacts emerging from motion would be of clinical value. We propose a supervised deep-learning based method to reduce motion artifacts in CBCT reconstructions.

Methods: Training and validation data with and without motion artifacts is generated using a patient-model based motion simulation. Input for the simulation are three training and one validation 4D-CT patient scans. Each has been deformed using six free breathing and eight breath-hold curves, the latter to simulate effects off residual motion. For augmentation purposes the imaging center of each simulation is shifted randomly (+-10mm). Input and output pairs are extracted from the reconstructed volume as overlaying patches with size 256x256x32 (2x2x2 mm3) such that a total of 5 patches is extracted for each volume. This results in a training set with 220 and a validation set with 50 examples. A V-net 3DCNN (kernel size 5x5x3) is trained to predict the artifact from the volume with artifacts. Artifact images are generated by subtracting the ground truth CBCT from the input CBCT. To demonstrate the generalization of the method we applied this method on 5 HalcyonTM breath-hold test data where no ground truth exists.

Results: The mean absolute validation error during training demonstrates the ability of the method to estimate the motion artifacts for previously unseen data. Visual inspection of the predicted artifact images on the validation and test data show reduced artifacts without noticeably affecting image quality and resolution.

Conclusion: To further improve these first results, we are planning to provide pre-processed volumes in additional input channels as well as larger training sets. Ensuring consistency with the raw data and quantification of the results (STD, entropy, MSE reduction) will be addressed in a next step.

Funding Support, Disclosures, and Conflict of Interest: All authors are full time employees of Varian Medical Systems


Cone-beam CT


Not Applicable / None Entered.

Contact Email