Deep CNN-Based Ultrasound Super-Resolution for High-Speed High-Resolution B-Mode Imaging

In real-time high-resolution B-mode ultrasound (US) imaging, the lateral resolution, or the number of scan lines, may be limited due to the speed of sound, if a longer penetration depth is needed as in obese patient imaging. To deal with this limitation, we propose to apply a super-resolution (SR) technique to US B-mode scan images of low lateral and high depth resolutions. Recently, several deep convolutional neural networks (CNNs) have shown good performance in SR of natural images. However, they usually provide rare improvement on image textures. To alleviate this problem, SRGAN is proposed. In the US image, speckle noise can be considered as a texture. It is furthermore observed that its degree of fineness is important to determine the image resolution, in addition to structural sharpness. We hence adopt and modify the SRGAN to make B-mode US images of low lateral resolution similar to their original high-resolution (HR) images. We first define a problem of US SR as the recovery of HR images from 4-to-l laterally sub-sampled low-resolution (LR) images. Those LR images are acquired by using a quarter of the number of scan lines required for HR images. LR images thereby have asymmetric resolution, or low lateral and high depth resolution. It should be noted in this acquisition that the beam width is widened four times with respect to the acquisition of HR images, to reduce unwanted aliasing artifacts along the lateral direction. In order to efficiently improve the lateral resolution, we slightly modify the network architecture of SRGAN.