Abstract:

The continuous development and widespread use of computed tomography(CT) in medical practice have caused patients and doctors’ attention to the risk of cancer due to radiation dose. However, spectral CT can generate severe Poisson noise as a result of the division of multiple energy bins while reducing the radiation dose, which has an adverse effect on clinical diagnosis. Aiming at improving the reconstructed image quality of spectral CT under low radiation dose, a denoising model based on a generative adversarial network with the hybrid perceptual loss was proposed. The method combines pixel space with the human-perceived feature space into generative adversarial network, and introduces the residual learning to reuse features from different levels to avoid the loss of important details that come from the increase in the number of network layers. Then the Wasserstein distance is chosen instead of Jensen-Shannon to improve the stability of training. The inputs are from multiple energy bins, so that the spatial correlation of the reconstructed image from a single energy bin and the energy correlation between different reconstructed images from multiple energy bins can be utilized to improve the visual sensitivity of low dose spectral CT images. The experiments were performed on the simulated phantom data sets from Duke University Medical Center. The results show that the proposed method increases the PSNR by about 5 dB, the structure similarity (SSIM) value by about 0.2, and the feature similarity (FSIM) value about by 0.06. Compared with current advanced low-dose CT denoising method, the hybrid loss function based generative and adversarial network achieves a better noise removal, while retaining the necessary details for diagnosis. Moreover, the spectral information also plays a positive role in the recovery of noise images, and the use of hybrid perceptual loss is superior to any single loss. Overall, the proposed generative adversarial network with a hybrid loss method can significantly improve the quality of low-dose spectral CT images.

Key words: information processing technology, low dose, generative adversarial network, residual learning