(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose:
Generative adversarial network (GAN) has emerged as a powerful tool for medical imaging. GAN-based algorithms with an end-to-end convolutional neural network (CNN) may produce high-frequency checkerboard-like artifacts in the reconstructed images. In this study, we aimed to develop a novel GAN algorithm based on Discrete Cosine Transform (DCT) for low dose CTs (LDCT) reconstruction.
Methods:
The difference between LDCT images and high dose CTs (HDCT) images mainly lies in the high-frequency region in the DCT domain. We proposed a novel architecture for the conditional Wasserstein GAN (c-WGAN) to emphasize the learning of high-frequency components for LDCT reconstruction. A weighted DCT-based loss function was designed for the GAN generator. LDCTs images are fed to the GAN generator that produces images close to those reconstructed from HDCT. The GAN discriminator was designed to distinguish between HDCT images and images produced from the generator. Both the generator and discriminator are based on the CNN structure. A weight matrix was generated according to the change rate quantified between training image pairs of LDCT and HDCT, which gives greater penalty to the high-frequency difference than the low-frequency difference.
Results:
We conducted experiments on the 2016 AAPM Low Dose CT Grand Challenge dataset, and evaluated LDCT reconstructions over HDCT images by Peak Signal-to-Noise Ratio (PSNR). Our method performed the best. Particularly, average PSNRs of FBP, TV, c-WGAN and ours were 29.25, 31.27, 32.14 and 32.93 respectively. In addition, our method also performed the best in suppressing artifacts and recovering image details.
Conclusion:
We demonstrated the efficiency of the DCT-based GAN method for the reconstruction of LDCTs, which is conducive to clinical medical diagnosis to recover better details for LDCT than other state-of-the-art methods.
Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by the National Natural Science Foundation of China (NNSFC), under Grant Nos. 61375018 and 61672253.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
