Highly efficient encoder-decoder network based on multi-scale edge enhancement and dense connectivity for LDCT image denoising

Aiming at the problem of low-dose computed tomography (LDCT) image quality degradation, this paper proposes a highly efficient encoder decoder CNN network(HEDCNN) based on multi-scale edge enhancement and dense connectivity. Firstly, The network uses the multi-scale feature learnable edge enhancement block to extract the edge information of the original image and fuses it into all the network model through dense connectivity, which enhances the ability of the network to recover the edge information. Secondly, this network uses hybrid dilated convolution to achieve multi-scale receptive field coverage and improve the model's ability to process details. Finally uses the multi-scale reconstruction block to help the network to reconstruct the recovered image. A compound loss that combines the Huber loss and the SSIM loss is introduced to overcome the problem of the model's over-smoothing, and to strengthen the network's denoising ability. The performance of the proposed method is validated and evaluated based on the LDCT Challenge Abdomen data set from the AAPM-Mayo clinic. Compared with state-of-the-art model designs in different architectures, HEDCNN achieves optimal metrics without a large number of learned parameters, 33.6653 for PSNR, 8.4334 for RMSE, and 0.9138 for SSIM. The experiments demonstrate that the designed HEDCNN is a excellent model for noise suppression and structure preservation.