Brain MR Image Super-resolution using 3D Feature Attention Network

Magnetic resonance images (MRI) with high spatial resolution provide detailed anatomical information for accurate disease diagnosis and quantitative analysis. However, the resolution of clinical MRI is restricted by hardware limitation and cost. Recently, convolutional neural networks (CNNs) are utilized to improve the spatial resolution of MRI. Whereas, current CNN-based super-resolution (SR) methods treat different types and levels of features equally, which hinders the representation ability of network. In this paper, we propose a novel feature attention super-resolution (FASR) network to adaptively capture different informative features. FASR uses parallel channel and spatial attention to enhance valuable features and suppress redundant information. The refined features are upsampled using a sub-pixel convolution layer, and then fused to predict the missing high-resolution details. To accelerate the training and generate realistic MRI, we introduce cross-scale residual and adversarial training to train FASR. Experimental results on brain MRI datasets of healthy subjects and gliomas show that the proposed FASR achieves a new state-of-the-art MRI SR performance.