Low-dose CT image processing based on controllable residual U-shaped network

Medical Computed Tomography(CT) requires the use of the lowest possible low-dose scanning mode. However, too low dose scanning will produce low-quality CT images containing a large amount of noise and artifacts affecting the diagnosis, and the intensity and distribution of noise and artifacts in different parts of the tomographic images under the same dose are not the same. For this reason, this paper adopts a dynamically controllable residual network to improve the CT image restoration effect, the main idea of this network is to add a control variable to the conventional residual network connection to allow the acquisition of the weighted sum of residuals, and realize the dynamic controllability of residuals. With the help of conditional residual blocks, a tunable UNet (DR-UNet) is designed in conjunction with the UNet. The network consists of two parts: the base network and the conditional network, which will consist of the conditional residual block, the downsampling layer and the upsampling layer, which consists of the encoder path and the decoder path. The conditional network takes the image quality assessment coefficients as inputs, and a modified form of multilayer fully connected network will be used to improve the accuracy of the control weight estimation. Experimental results show that the proposed method exhibits competitive performance in terms of noise suppression, structural realism and visual browsing improvement.