Click here to

Session: Advances in CT II [Return to Session]

Fast and Motion Robust CT Super Resolution Using Sinogram

y yoon1,2*, J Chun1,2,5, J Baek3, J Choi4, J Kim1,2,5**, (1) Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, KR (2) Medical Physics and Biomedical Engineering Lab (MPBEL), Yonsei University College of Medicine, Seoul, KR, (3) School of Integrated Technology, Yonsei University, Incheon, KR, (4) College of Engineering, Ewha Womans University, Seoul, KR,(5) Oncosoft, Seoul, KR


TU-D930-IePD-F8-5 (Tuesday, 7/12/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 8

Purpose: To observe as low as reasonably achievable(ALARA) principle, many studies proposed x-ray reduction methods by super resolution(SR) on sparse view CT or detector binning in low dose using both sinogram and reconstruction image. However, additional processing time required for image reconstruction and patient motion artifacts are unsolved remaining challenges. In this study, we propose fast deep learning based sinogram SR to restore resolution of binned CT regardless of motion artifact using only sinogram.

Methods: In this study, we simulated detector binning by downsampling sinogram as the factor of 2. Proposed SR model trained and evaluated in sinogram acquired from 30 and 25 (13,438 and 11,483 slices) XCAT head & neck phantoms and forward projected 12 and 8 (1,200 and 800 slices) real patients from CQ500 dataset. Moreover, with aforementioned XCAT datasets, artificial motion applied to compare robustness of sinogram based SR and image based SR

Results: In XCAT sinogram SR in the factor of 2, mean absolute error(MAE) of image was 2.95e-4, 1.78e-4, 1.20e-4 without motion and 2.78e-4, 1.77e-4, 1.18e-4 with motion and in CQ500 sinogram SR, peak signal to noise ratio(PSNR) of image was 60.4, 61.7, 63.1 in bi-cubic interpolation, SRGAN, proposed SR, respectively. In motion datasets, MAE of image SR with same proposed model showed 2.60e-4 in without motion, whereas 3.76e-4 in with motion. Average inference time of proposed SR takes 122±4.8ms for a slice. Which is shorter than sinogram filtered back projection(FBP) takes 145±13.4ms.

Conclusion: In this study, we demonstrate that proposed model have fast and successfully restored resolution of binned sinograms in the factor of 2, regardless of motion artifact by using only sinogram. We expect the proposed motion robust SR model to be especially useful in reducing x-ray exposure in situations with detector binning.

Funding Support, Disclosures, and Conflict of Interest: Funding: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2020R1A4A1016619). Conflict of interest: Jin sung kim declared that he is co-founder of oncosoft.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email