Transformer for Ultrafast Ultrasound Localization Microscopy

Ultrasound localization microscopy (ULM), which allows for the detection of subwavelength acoustic scattering of intravenously injected microbubbles, enabling non-invasive visualization of microvessels on a micron scale. Unfortunately, the implementation of ultrafast ULM is still constrained by the trade-off between data acquisition time, data processing efficiency and image spatial resolution. More recently, deep learning-based methods (DL-ULM) have exhibited promising potential for accelerating ULM imaging by reducing the number of ultrasound (US) image data (
$L$
frames) required due to the ability to localize to as many microbubbles as possible. Nevertheless, this requires the accumulation of a certain number of microbubble events to obtain the desired super resolution image, which results in data acquisition times of typically tens of seconds. To facilitate ultrafast ULM imaging, in this article, we present a novel netwok, termed as Trans-ULM. Specifically, a multi-branch CNN combined with a recursive transformer to reconstruct a super-resolved image directly from a temporal mean low-resolution image generated by averaging
$l$
-frame raw US images with
$l$
being considerably smaller than
$L$
. The experimental results indicate that, trans-ULM can significantly reduce imaging time as compared to the recently reported ULM-DL method, while maintaining spatial resolution.