Résumé: In order to solve the problems of low SNR and low use value of low-dose CT images, this study proposes a multi- layer enhancement of low-dose CT images via adaptive fusion. In our study, the image denoising training based on generative adversarial network is carried out, and the perceptual loss and structural loss optimization generator are used to strengthen the denoising ability and retain the details of the image. To clearly observe the pathological tissue structure, it is necessary to perform a certain degree of image enhancement and image fusion on CT images. Using the real clinical data disclosed in the AAPM competition as the experimental dataset, in the image denoising experiment, the PSNR, SSIM, and RMSE are 33.0155, 0.9185, and 5.99. Compared to traditional methods, the effectiveness of the proposed method was better by 10.76%, 4.08% and 24.54% on average, respectively. The proposed model in this study obviously reduces the noise of CT images, and the obtained CT images are more detailed, its brightness and contrast are significantly enhanced, which proves the feasibility and effectiveness of the algorithm.

Langue: Anglais