Learning-based super-resolution of 3D ultrasound images

The super-resolution (SR) of 3D ultrasound (US) images reconstructs and increases the number of slices, maintaining the high-resolution (HR) image visually like the corresponding low-resolution (LR) while improving anatomical structures' visibility and avoiding artefacts. We propose a novel learning-based super-resolution method of 3D US images that accounts for the presence of speckle noise and preserves the properties of the high-resolution image in terms of texture patterns and visual appearance. We specialise our model to speckle noise and 2X and 4X up-sampling factors. We apply the trained network to reconstruct non-acquired slices of 3D US images from the forearm anatomical district of an open US data set. Finally, we evaluate the quantitative results compared to previous work according to feature, local, and noise-based metrics.