A U-Residual Local Window Transformer for Ultrasound Images Super-Resolution Reconstruction

Ultrasound images have inherent disadvantages, such as blurred details and high noise, and super-resolution (SR) reconstruction will improve the image quality and help for more accurate clinical diagnosis. In order to solve the problem that it is difficult to capture local feature information and lose important information in the process of downsampling, we propose an SR reconstruction method of ultrasound single image based on U-residual local window transformer (U-RLWT). Specifically, we construct a network encoder module that fuses two residual swim transformer blocks and a downsampling layer. It obtains more low-frequency information from ultrasound images by improving the receptive field of the feature extractor. Then, to combine the feature information between the layers and recover the feature map, the decoder module was constructed by using the residual swim transformer block and the upsampling layer. Finally, the reconstruction module was designed with convolution and Pixel-shuffle network to realize the SR reconstruction of ultrasound single images. Experimental results show that, compared with the advanced reconstruction methods, the proposed method can reconstruct the better results compared with the ground truth. When applied to larynx ultrasound images and PICMUS datasets, it achieves the highest peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) on the scale factors of $\times 2$ and $\times 4$ .