Enhanced fully convolutional network based on external attention for low-dose CT denoising

In recent decades, researchers have concentrated their efforts on developing techniques to denoise low-dose, sparse-angle CT images. Currently, mainstream low-dose CT methods include traditional iterative reconstruction algorithms and deep learning algorithms. However, these methods still exhibit transition smoothing, noise, and artifact residue. Inspired by the idea of deep learning, we propose an improved fully convolutional neural network using external attention. The network incorporates a novel external attention module, and use perceptual loss as the loss function which is used to construct a fully convolutional model with dense connections. This model fuses extracted feature information to achieve image denoising. Following training, the proposed EA-FCN demonstrated superior performance on actual clinical CT images. Furthermore, our method achieved favorable subjective and objective evaluations in terms of noise and artifact suppression.