Texture Transformer Super-Resolution (TTSR) for Patient CT Images
S Zhou1*, L Yu2, M Jin1, (1) University of Texas at Arlington, Arlington, TX, (2) Mayo Clinic, Rochester, MN
Presentations
TH-D-207-6 (Thursday, 7/14/2022) 11:00 AM - 12:00 PM [Eastern Time (GMT-4)]
Room 207
Purpose: It was previously demonstrated texture transformer image super-resolution (TTSR) can improve the spatial resolution and suppress noise of the low-dose computed tomography (CT) images using the simulated data. In this work, we further developed TTSR for patient CT images and compared it with another deep-learning super-resolution (SR) method.
Methods: TTSR was previously developed for CT SR, which is a reference-based deep-learning method and can transfer the low-quality CT images to the high-quality CT images using simulated XCAT phantom data. In this work, we aim to apply and evaluate TTSR for the real patient data. First, we iteratively reconstruct high-quality (full-dose, 512x512 resolution) and low-quality (quarter dose, 128x128 resolution) CT images using the projection data from the 2016 NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge. Then, the TTSR models are trained either using the pre-trained model or from scratch using nine patients’ images to establish the mapping from low-quality images to high-quality images. Finally, the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are used as the quantitative metrics using the remaining one patient’s images as the test set. For comparison, quarter-dose full-resolution reconstruction (with 4x data) and the generative adversarial network (GAN) with cycle consistency (GAN-CIRCLE) are used.
Results: The PSNR and SSIM performance of different methods is ranked from the worst to the best: quarter-dose full-resolution (PSNR=22.13±1.08dB and SSIM=0.39±0.04), GAN-CIRCLE (PSNR=27.15±0.53dB and SSIM=0.53±0.04), TTSR pre-trained (PSNR=27.56±0.59dB and SSIM=0.56±0.05), and TTSR from scratch (PSNR = 30.31±1.37dB and SSIM=0.68±0.06). Quarter-dose full-resolution images suffer great noise. TTSR provides much more structural details than GAN-CIRCLE. The artifacts introduced by TTSR using the pre-trained model can be alleviated by TTSR from scratch.
Conclusion: For low-quality real patient data, TTSR trained from scratch can provide much more improved super-resolution CT images than GAN-CIRCLE.
Funding Support, Disclosures, and Conflict of Interest: This work is supported in part by the U.S. National Institutes of Health under Grant No. 1R15HL150708-01A1.
Keywords
Low-dose CT, Noise Reduction, Spatial Resolution
Taxonomy
IM- CT: Machine learning, computer vision
Contact Email