| Clinical diagnosis and quantitative analysis usually require high-resolution magnetic resonance(MR)images.Newborns have small brains,immature brains,and lack of dedicated imaging equipment.Neonatal images collected by MRI scanners have distortions such as low resolution,low signal-to-noise ratio,and partial volume effects.The superresolution(SR)technology is an effective and reliable method to improve the resolution of neonatal magnetic resonance images.Compared with the traditional image super-resolution method,the deep learning-based method can obtain higher quality magnetic resonance images.In this paper,a joint network that can be reconstructed simultaneously with denoise and super-resolution is proposed to improve the resolution of neonatal magnetic resonance images.The main work is as follows:(1)First of all,the attention mechanism and the dual-domain learning strategy are combined to propose a neonatal 3D magnetic resonance image super-resolution reconstruction algorithm based on the dual-domain attention network.Different from the traditional attention mechanism,this paper not only calculates the channel attention of the feature map,but also calculates the spatial attention of the feature map to enhance the expression ability of the network to the feature.In this paper,the 3D-MR image is converted from image domain to k space by discrete cosine transformation(DCT),and then the inverse transformation of discrete cosine transformation is used to transform this feature information to image domain,which is fused with the feature information extracted from the image domain,and improves the learning ability of the network to feature.The experimental results show that the attention mechanism and dual-domain learning strategy make the network performance greatly improved,and the quality of the reconstruction image is obviously better than that of other methods.(2)Based on the super-resolution network proposed in this paper,we design a joint network for denoising and super-resolution reconstruction of 3D-MR images.The joint network first denoises the input image,then maps the feature map that removes the noise,and finally obtains the super-resolution 3D-MR image.During the joint network training,we combine noise loss and super-resolution loss to update the network.The experimental results show that the joint network performance of denoising and super-resolution is better than that of other task-ordered joint networks,and also better than the method of performing two tasks separately.In short,the method proposed in this paper has the best performance. |
PDF Full Text Request