A Spatiotemporal Denoising Method for Low-Dose Cardiac CT Images
J Yang1, S Zhou1*, J Huang1, L Yu2, M Jin1, (1) University of Texas at Arlington, Arlington, TX, (2) Mayo Clinic, Rochester, MN
Presentations
TU-IePD-TRACK 1-3 (Tuesday, 7/27/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Purpose: To develop a spatiotemporal denoising method using an unpaired video-to-video translation model, which can improve the image quality of low-dose cardiac CT images by exploiting both spatial and temporal correlations of dynamic image frames.
Methods: The new spatiotemporal denoising model is built upon the generative adversarial network with cycle consistency (CycleGAN), which is an unpaired image-to-image translation model. In order to utilize the temporal dynamics of image frames, a recurrent loss of image frames in both domain (low-dose and full-dose CT images) is added to enforce the temporal correlation between image time frames. This is particularly useful for denoising of low-dose dynamic CT images, such as multi-phase cardiac CT angiography, where consecutive cardiac gates exhibit strong temporal dependencies. The 4D XCAT phantom program based on 18 patients’ data is used to generate both benchmark full-dose and low-dose (80% reduction) cardiac CT images (8 gates for each patient). 17 of 18 patients are used as training data and the rest one is used as test data. To qualitatively evaluate the model performance, peak signal noise ratio (PSNR) and structural similarity index (SSIM) are used.
Results: The spatiotemporal denoising method achieves better quantitative performance (PSNR: 38.35 ± 1.46 dB and SSIM: 0.9300 ± 0.0112) than the low-dose images (PSNR: 31.56 ± 1.74 dB and SSIM: 0.8436 ± 0.0284) and CycleGAN (PSNR: 36.85 ± 1.29 dB and SSIM: 0.9240 ± 0.0122). It also significantly improves the low-dose CT image quality compared to CycleGAN.
Conclusion: We proposed a new unpaired deep-learning based denoising method using spatial cycle consistent and temporal recurrent consistency for low-dose CT image sequences. The simulation study demonstrated the superior performance of the proposed method. The future study will investigate its effectiveness on real patient data.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by the U.S. National Institutes of Health under Grant No. 1R15HL150708-01A1. No conflict of interest.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email