Attention USR-Net: An End-to-End Mapped Ultrasound Localization Microscopy

Ultrasound localization microscopy (ULM) breaks the wave diffraction limit for super-resolution imaging of microvessels deep within tissues. However, it remains challenging since traditional ULM requires a sufficient accumulation of microbubbles over a long period to accurately delineate the microvasculature. Furthermore, elaborate microbubble localization is required for the mapping from ultrasound image sequence to super-resolution image. To achieve ultrafast ULM, we propose an Attention U-shaped Super-Resolution Net (USR-Net) to directly map raw ultrasound sequences to super-resolution images by end-to-end deep learning. Specifically, USR-Net integrates the U-Net architecture and attention mechanisms to significantly shorten imaging time and improve the quality of super-resolution images. We validate our network on public datasets. Extensive experiments demonstrate the superiority of USR-Net in fast super-resolution imaging, showing higher SSIM and PSNR when compared with existing deep learning-based fast ULM method. Particularly, USR-Net reduces the processing time by a factor of 7 times when compared with conventional ULM technology.