Super-resolution of medical image using representation learning

Super-resolution (SR) of single image is a meaningful challenge in medical images based diagnosis, while the image resolution is limited. Also, numerous deep neural networks based models were proposed and achieve excellent performance which is superior to the previous handcrafted methods. In this paper, we employ a deep convolutional neural networks for the super-resolution (SR) of single medical image, which learns the nonlinear mapping from the low-resolution space to high-resolution space directly. In addition, we use three sets imaging data (Mammary gland, Prostate tissue and Human brain) training deep network respectively. Firstly, we use Randomized Rectified Linear Unit (RReLU), which incorporates a nonzero slope for negative part to solve the problem of over compression. Secondly, for the purpose of enhancing the quality of reconstructed result and reducing the noise of over-fitting, Nesterov's Accelerated Gradient (NAG) method on the SRCNN is used to accelerate the convergence of loss function and avoid the large oscillations. A comparative performance evaluation is carried out over a set of experiments using real imaging data to verify the validity of proposed algorithm.