Low-dose CT image denoising based on residual attention mechanism

There are complex image noises in LDCT images. Among the existing deep learning models, although there are many excellent models that can restrain the noise of LDCT images, the denoised images of these models will produce a smooth effect and some will be lost. Image details. To improve the image quality and strengthen the correlation between feature channels, inspired by the REDCNN network and CBAM module, this paper proposes an improved REDCNN network (CASA-REDCNN) based on the attention mechanism. By adding channel attention and spatial attention to the residual connection part of REDCNN, the attention module can distinguish the differences between feature channels, focus more attention on channels containing noise and artifacts, and reduce attention on other channels. force. Experiments show that this network can effectively reduce noise and artifacts in LDCT images, and can retain more image details in key areas. It also has good denoising capabilities for LDCT images with different noise levels, proving that this network has good denoising ability and improves the performance of the original model.