An Application of Super-Resolution Generative Adversary Networks for Quasi-Static Ultrasound Strain Elastography: A Feasibility Study

In this work, a super-resolution approach based on generative adversary network (GAN) was used to interpolate (up-sample) ultrasound radio-frequency (RF) echo data along the lateral (perpendicular to the acoustic beam direction) direction before motion estimation. Our primary objective was to investigate the feasibility of using a GAN-based super-solution approach to improve lateral resolution in the RF data as a means of improving strain image quality in quasi-static ultrasound strain elastography (QUSE). Unlike natural scene photographs, axial (parallel to the acoustic beam direction) resolution is significantly higher than that of lateral resolution in ultrasound RF data. To better handle RF data, we first modified a super-resolution generative adversary network (SRGAN) model developed by the computer vision community. We named the modified SRGAN model as super-resolution radio-frequency neural network (SRRFNN) model. Our preliminary experiments showed that, compared with axial strain elastograms obtained using the original ultrasound RF data, axial strain elastograms using ultrasound RF data up-sampled by the proposed SRRFNN model were improved. Based on the Wilcoxon rank-sum tests, such improvements were statistically significant ($p <; 0.05$ ) for large deformation (3-5%). Also, the proposed SRRFNN model outperformed a commonly-used method (i.e. bi-cubic interpolation used in MATLAB [Mathworks Inc., MA, USA]) in terms of improving axial strain elastograms. We concluded that applying the proposed (SRRFNN) model was feasible and good-quality strain elastography data could be obtained in in vivo tumor-bearing breast ultrasound data.