A deep convolutional network for medical image super-resolution

This paper presents a novel deep network model specifically for medical image super-resolution reconstruction. We take full advantage of the fact that all medical images basically have distinct repetitive structure and large black border without any texture information. Compared with the Super-Resolution Convolution Neural Network (SRCNN) structure, firstly, we add a convolution layer to carry out secondary feature extraction which make the feature more representative. Then we add overlapping pooling layers in order to highlight the important features and fine processing them. At last, in order to use local features and global features to complete the reconstruction together, we add a link layer, establish a connection between the second pooling layer and the reconstruction layer. The experimental results show that average PSNR gains achieved by our algorithm are higher than the original SRCNN. The reconstructed CT images can clearly provide important reference for clinicians to make the correct treatment decisions, and also have important guiding significance for the difficulty and risk assessment of surgical feasibility.