Fast super-resolution ultrasound microvessel imaging using spatiotemporal data with deep fully convolutional neural...

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Ultrasound localization microscopy (ULM) offers sub-diffraction resolution for microvasculature imaging but suffers from long acquisition and processing times.
• Current deep learning ULM (deep-ULM) methods primarily use spatial information, neglecting the rich temporal data from microbubbles (MBs).

Purpose of the Study:

• To develop a deep neural network that leverages both spatial and temporal information for improved MB localization in deep-ULM.
• To enhance the resolution and reduce processing time for super-resolved ultrasound microvasculature imaging.

Main Methods:

• A deep neural network was trained using spatiotemporal ultrasound datasets simulating realistic MB signals.
• The network was validated on chicken embryo chorioallantoic membrane and in vivo human liver datasets, comparing results with optical microscopy and power Doppler imaging.

Main Results:

• The proposed method significantly improved spatial resolution, reducing microvessel full-width-half-maximum (FWHM) from 133 μm to 35 μm in chicken embryos.
• In vivo human liver imaging resolved microvessels with 170 μm FWHM and separated adjacent vessels 670 μm apart, outperforming power Doppler.
• Processing time for a high-resolution frame was approximately 16 ms, comparable to existing deep-ULM techniques.

Conclusions:

• Integrating spatiotemporal information into deep learning models substantially improves MB signal identification and ULM performance.
• This novel deep-ULM approach achieves superior resolution and contrast for microvasculature imaging with efficient processing times.
• The method holds promise for advanced non-invasive vascular imaging in clinical and research settings.