Deep learning-based fast and dense microbubble localization for ultrasound localization microscopy

Abstract

Ultrasound localization microscopy (ULM) has gained substantial attention owing to its ability to super-resolve minute blood vessels at clinically relevant imaging depth. However, accurate localization of individual microbubbles (MBs) in areas with high MB concentration and overlapping point spread functions (PSFs) remains a challenge. Furthermore, existing localization methods based on pre-determined MB PSFs cannot reflect the highly non-stationary PSFs that vary spatially with MB concentration and nonlinear responses, ultrasound diffraction, and imaging settings. To address these limitations, we implemented the DECODE (DEep COntext Dependent) neural network that was recently developed for optical imaging on ULM. DECODE constructs a Gaussian mixture model with a mixed MB count and localization loss to output the probability, uncertainty, and sub-pixel location corresponding to each MB detection, achieving accurate identification and localization of MBs. Notably, DECODE was trained with realistic simulation data that incorporates MB brightness, movement, ultrasound system noise, and PSFs produced by a generative adversarial network that encodes the internal distribution of MB PSFs obtained from in vivo ultrasound imaging. In high MB-density regime, simulation studies demonstrated that DECODE improved MB detection rate from 41% to 95%, and reduced localization error from 109.4 μm to 32.5 μm (20 MHz) when comparing to conventional MB localization techniques. DECODE also demonstrated improved in vivo ULM imaging in mouse brain.