Pruning Sparse Tensor Neural Networks Enables Deep Learning for 3D Ultrasound Localization Microscopy.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Computational Biology

Background:

• Ultrasound Localization Microscopy (ULM) images micro-vessels in vivo with micron resolution.
• Current ULM requires long acquisition times or advanced algorithms for high microbubble concentrations.
• Existing deep learning methods for ULM are limited to 2D due to memory constraints.

Purpose of the Study:

• To develop a deep learning-based 3D ULM method using sparse tensor neural networks.
• To improve memory scalability for high-dimensional ultrasound data.
• To enable faster and more detailed in vivo micro-vessel imaging.

Main Methods:

• Utilized sparse tensor neural networks for 3D ULM.
• Investigated methods for converting ultrasound data to a sparse format.
• Evaluated the impact of data sparsity on information loss and performance.

Main Results:

• Sparse formulation reduced memory requirements by 2x in 2D ULM with minimal performance loss.
• In 3D ULM, sparse tensor networks decreased memory usage by two orders of magnitude.
• The 3D approach outperformed conventional ULM in high microbubble concentration settings.

Conclusions:

• Sparse tensor neural networks effectively enable 3D ULM, overcoming memory limitations of dense networks.
• This method allows for higher microbubble concentrations and reduced acquisition times, similar to 2D deep learning ULM.
• The developed 3D ULM technique offers significant improvements for in vivo micro-vessel imaging.