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

Aiming at the problems of edge loss due to down-sampling-up-sampling operations of encoder decoder networks and the difficulty of extracting global features from networks due to the difficulty of obtaining a large receptive field for traditional convolution, a highly efficient encoder decoder convolutional neural network HEDCNN based on multi-scale edge enhancement and dilated convolution is proposed. HEDCNN uses a multi-scale feature learnable edge enhancement block to extract the multi-scale edge information of the original image and fuses it into the network through dense connections, improving the network's ability to recover edge information. Hybrid dilated convolution expands the receptive field of the network, allowing the network to capture broader contextual information while also focusing on local detail information. A compound loss function combining Huber loss and SSIM loss overcomes the over-smoothing problem of the model and further preserves the details of the denoised image. In the 2016 NIH AAPM-Mayo Clinic Low-Dose CT Challenge data sets, compared with other state-of-the-art LDCT denoising models, HEDCNN obtains the best PSNR and SSIM. At the same time, the denoising image is closest to the NDCT image in terms of visual effect.