Unsupervised Super-Resolution Framework for Medical Ultrasound Images Using Dilated Convolutional Neural Networks

Abstract

Ultrasound Imaging is one of the most widely used imaging modalities for clinic diagnosis, but suffers from a low resolution due to the intrinsic physical flaws. In this paper, we present a novel unsupervised super-resolution (USSR) framework to solve the single image super-resolution (SR) problem in ultrasound images which lack of training examples. Our method utilizes the powerful nonlinear mapping ability of convolutional neural networks (CNNs), without relying on prior training or any external data. We exploit the multi-scale contextual information extracted from the test image itself to train an image-specific network at test time. We utilize several techniques to improve the convergence and accuracy, including dilated convolution and residual learning. To capture valuable internal information, dilated convolution is employed to increase the receptive field without increasing the network parameters. To speed up the convergence of the training, residual learning is used to directly learn the difference between the high-resolution and low-resolution images. Quantitative and qualitative evaluations on real ultrasound images demonstrate that the proposed method outperforms the state-of-the-art unsupervised method.