Deep Learning-Enhanced Super-Resolution Ultrasound for Efficient Brain Vascular Mapping

Robotic implantation of neural electrodes for in-vasive brain-computer interface(BCI) requires precise vascular mapping to avoid major blood vessels. While super-resolution ultrasound imaging offers superior resolution compared to traditional methods, it typically requires probes with numer-ous elements, demanding substantial hardware resources. This paper presents a novel approach combining deep learning with reduced-element probes for high-resolution vascular imaging. The proposed method employs a transformer-based neural network to accurately localize high-density microbubbles while reducing the required number of probe elements. Simulations reveal that a 32-element linear probe with deep learning outper-forms a 64-element probe using 2D normalized cross-correlation for microbubble concentrations below 1.95 microbubbles/mm2. The overall localization precision and Jaccard index demon-strate superior performance of the deep learning approach, particularly at lower microbubble densities. In vivo experiments on a mouse brain model using a 64-element probe (15 MHz) and deep learning achieved vascular network reconstruction with 21 um resolution, distinguishing vessels 26 um apart. This technique significantly lowers the hardware requirements for constructing high-resolution vascular networks, thereby providing detailed cerebrovascular distribution maps. These maps have the potential to enable more precise planning and safer execution of neural electrode implantation for BCI research, while reducing overall system complexity.