Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Artifacts Reduction of CBCT Based On Self-Supervised Learning

G Liugang1,2*, X Kai1, J Sun1, C Qian1, H Bi1, N Xinye1, (1) The Affiliated Changzhou NO.2 People's Hospital of Nanjing Medical University, Changzhou, Jiangsu, CN,(2) School of computer science and engineering,Southeast University, Nanjing, Jiangsu, CN,


PO-GePV-M-316 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: In this study, we proposed an artifacts reduction network based on self-supervised learning, which focuses on removing severe streak artifacts in CBCT-to-CT generation.

Methods: Three hundred patients undergoing radiotherapy were included in this study. The CBCT and planning CT images were rigid registered and approved by senior oncologists. The images of 270 patients were used as training data and remained 30 patients as testing data. The proposed self-supervised streak artifacts reduction network (SARN) were trained only using planning CT images. The cycle-consistent generative adversarial networks (cycleGAN) were trained in an unpaired manner to generate synthetic CT (sCT). The trained SARN was tested combined with cycleGAN, the sCTs generated from cycleGAN were input to SARN to obtain the final sCT (cycleGAN-SARN), or CBCT images were input to SARN and output were token as cycleGAN’s input to generate final sCT (SARN-cycleGAN). The trained models were evaluated using testing data and generated sCT images were compared.

Results: At the site of thorax, mean absolute error (MAE) between CT and sCT was 43.4±6.9 HU, 42.7±6.9 HU and 42.8±6.7 HU for cycleGAN, cycleGAN-SARN and SARN-cycleGAN respectively, comparing to 71.6±19.3 HU between CT and original CBCT. At the site of abdomen, MAE of CBCT was 53.2±9.6 HU, reducing to 31.8±5.4 HU, 30.0±4.9 HU and 30.3±4.8 HU of sCTs for cycleGAN, cycleGAN-SARN and SARN-cycleGAN respectively. In the comparison of MAE of sCT images generated by cycleGAN, cycleGAN-SARN and SARN-cycleGAN, significant differences (P<0.05) were found at thorax and abdomen.

Conclusion: The proposed SARN reduced severe streak artifacts in CBCT-to-sCT generation, especially at the site of abdomen and thorax. The combination of SARN and cycleGAN can remove more artifacts and improve the quality of generated sCT images.

Funding Support, Disclosures, and Conflict of Interest: This work is supported by Changzhou Sci&Tech Program (No.CJ20200099 and CJ20210128), General Program of Jiangsu Provincial Health Commission (No. M2020006), Changzhou Key Laboratory of Medical Physics (No. CM20193005).


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email