Purpose: To improve the spatial resolution and to suppress the noise of the low-dose computed tomography (CT) images, we adapted a Texture Transformer network for image Super-Resolution (TTSR) method.
Methods: The TTSR is a reference-based image super-resolution method. We adapted it to translate the low-dose (low-resolution and noisy) CT images to the full-dose (high-resolution and clean) CT images. The low-dose CT images and full-dose images are severed as queries and keys in a transformer, respectively. The image translation is optimized through deep neural network (DNN) texture extraction, relevance embedding, and attention-based texture transfer and synthesis to enable joint feature learning between low-dose and full-dose images. 4D XCAT phantom program based on 18 patients’ data was used to generate both benchmark full-dose CT images (512x512 in-plane resolution) and low-dose images (20% dose and 128x128 in-plane resolution). The data from 17 patients were used as training data and the rest one was used as test data. The reference images in TTSR were randomly selected from the training set, which were not the same as the test images. The peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were used as the quantitative metrics. For comparison, we used cubic spline interpolation and generative adversarial network (GAN) with cycle consistency (GAN-CIRCLE).
Results: The performance of different methods (mean ± std) are PSNR = 31.57±0.89dB, and SSIM = 0.74±0.02 for cubic spline interpolation, PSNR = 34.85±0.32dB and SSIM = 0.88±0.02 for GAN-CIRCLE, and PSNR = 37.23±0.42dB and SSIM=0.84±0.02 for TTSR. TTSR also recovers more details in enhanced low-dose CT images than the other two methods, although some artifacts are noticeable.
Conclusion: TTSR based on texture transformer and attention mechanism is effective to improve the spatial resolution and to suppress the noise of low-dose CT images in general. Further work is needed to minimize the artifacts.
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.
Not Applicable / None Entered.