Recently, many studies have shown that deep convolutional neural network can achieve superior performance in image super resolution (SR). The majority of current CNN-based SR methods tend to use deeper architecture to get excellent performance. However, with the growing depth and width of network, the hierarchical features from low-resolution (LR) images cannot be exploited effectively. On the other hand, most models lack the ability of discriminating different types of information and treating them equally, which results in limiting the representational capacity of the models. In this study, we propose the multi-attention residual network (MARN) to address these problems. Specifically, we propose a new multi-attention residual block (MARB), which is composed of attention mechanism and multi-scale residual network. At the beginning of each residual block, the channel importance of image features is adaptively recalibrated by attention mechanism. Then, we utilize convolutional kernels of different sizes to adaptively extract the multi-attention features on different scales. At the end of blocks, local multi-attention features fusion is applied to get more effective hierarchical features. After obtaining the outputs of each MARB, global hierarchical feature fusion jointly fuses all hierarchical features for reconstructing images. Our extensive experiments show that our model outperforms most of the state-of-the-art methods.

Background: Brain computed tomography (CT) image registration is an essential step in the image evaluation of acute cerebrovascular disease (ACVD). Due to the complexity of human brain morphology, low brain CT soft-tissue resolution, low gray/white matter contrast, and the large anatomy variation across individuals, it is still a great challenge to perform brain CT registration accurately and rapidly. This study developed a hybrid supervised convolutional neural network (HSCN-Net) which may be used for assessment of ACVD in brain CT.Method: HSCN-Net generates synthetic deformation fields by a simulator to solve the lack of registration gold standard. The simulator are used to generate multi-scale deformation fields to overcome the registration challenge of large deformation. HSCN-Net adopts a hybrid loss function constituted by deformation field and image similarity to improve registration accuracy and generalization ability. In this work, one hundred and one brain CT images were included for HSCN-Net training and evaluation, and the results were compared with Demons and VoxelMorph. Qualitative analysis by visual evaluation, as well as quantitative analysis by Endpoint Error (EPE) between deformation fields, image Normalized Mutual Information (NMI), and Dice coefficient were carried out to access the model performance.Results: Qualitative analysis of HSCN-Net was similar to that of Demons, and both were superior to that of VoxelMorph. Moreover, HSCN-Net was more competent for large and smooth deformations. For quantitative evaluation, the EPE mean of HSCN-Net (3.29 mm) was lower than that of Demons (3.47 mm) and VoxelMorph (5.12 mm); the Dice mean of HSCN-Net was 0.96, which was better than that of Demons (0.90) and VoxelMorph (0.87); and the NMI mean of HSCN-Net (0.83) was slightly lower than that of Demons(0.84) but higher than that of VoxelMorph (0.81). In addition, the mean registration time of HSCN-Net (17.86 s) was lower than that of VoxelMorph (18.53 s) and Demons (147.21 s).Conclusion: The proposed hybrid supervised convolution registration network can achieve accurate and rapid brain CT registration. It is helpful for improving image evaluation of ACVD, thereby assisting clinicians in diagnosis and treatment decision-making.

No abstract

No abstract