Context-Aware Deep Learning Enables High-Efficacy Localization of High Concentration Microbubbles for Super-Resolution Ultrasound Localization Microscopy

Abstract

Ultrasound localization microscopy (ULM) is an emerging super-resolution imaging technique for deep tissue microvascular imaging. However, conventional localization methods are constrained by low microbubble (MB) concentration, as accurate localization requires a strict separation of MB point spread functions (PSFs). Furthermore, deep learning-based localization techniques are often limited in their ability to generalize to in vivo ultrasound data due to challenges in accurately modeling highly variable MB PSF distributions and ultrasound imaging conditions. To address these limitations, we propose a novel deep learning-pipeline, LOcalization with Context Awareness (LOCA)-ULM, which employs simulation that incorporates MB context to generate synthetic data that closely resemble real MB signals, and a loss function that considers both MB count and localization loss. In in silico experiments, LOCA-ULM outperformed conventional localization with superior MB detection accuracy (94.0% vs. 74.9%) and a significantly lower MB missing rate (13.2% vs 74.8%). In vivo, LOCA-ULM achieved up to three-fold increase in MB localization efficiency and a x 9.5 faster vessel saturation rate than conventional ULM.