A Deep Learning Framework for Spatiotemporal Ultrasound Localization Microscopy.

Main Methods:

• A three-dimensional convolutional neural network (CNN) based on a V-net architecture was trained using simulated ultrasound data of a mouse brain microvascular network.
• The CNN was trained to track microbubbles by learning from simulated datasets representing multiple microbubbles flowing through the vascular network.
• The deep learning approach was validated both in silico and in vivo using rat brain imaging.

Main Results:

• The deep learning approach demonstrated higher precision (81%) in reconstructing vascular networks compared to conventional ULM frameworks (70%) in silico.
• In vivo validation showed the CNN could resolve microvessels as small as 10 micrometers.
• The CNN approach offered improved resolution in vivo compared to conventional methods.

Conclusions:

• Deep learning significantly enhances the ability to recover dense vascular networks from ultrasound data with high microbubble concentrations.
• This novel method overcomes key limitations of traditional ULM, enabling faster and more precise microvascular imaging.
• The validated deep learning framework holds promise for advancing microvascular research and clinical applications.