Deep learning-based microbubble localization towards improved super-resolution ultrasound

Ultrasound (US) is an indispensable tool for visualizing the microvasculature noninvasively. Over the last decade, US localization microscopy (ULM), which exploits microbubble contrast agents as effective point scatterers, has dramatically improved the spatial resolution attainable with US subject to the so-called diffraction limit (millimeter scale), and enabled mapping of vessels on the order of 10 μm. However, microbubble localization, a principal component of ULM, typically relies on bubble sparsity and compounding of many (102 to 104) frames. Thus, to enable sensitivity to more transient phenomena at the smallest scales, optimally sensitive and specific means of localization are required. Drawing on newly available ground truth data (ULTRA-SR Challenge, IUS 2022), we apply convolutional neural network (DeepLabv3) to perform localization via segmentation of the microbubble regions and identification of their centroids. Following training on 80% of the available simulated US frames, the network demonstrated localization precision 0.88 and recall 0.50 (tolerance of 0.5λ) on the test frames, with bias on the order of 0.1λ, and inference time of 40 ms on Nvidia RTX3090. Such accurate and sensitive localizations have significant promise toward elucidating hemodynamics at still finer spatial and temporal resolutions.