Deep learning models for fast ultrasound localization microscopy

Ultrasound localization microscopy (ULM) can surpass the resolution limit of conventional ultrasound imaging. However, a trade-off between resolution and data acquisition time is introduced. For microbubble (MB) localization, centroid detection is commonly used. Therefore, low-concentrations of MBs are required to avoid overlapping point spread functions (PSFs), leading to a long data acquisition time due to the limited number of detectable MBs in an image frame. Recently, deep learning-based MB localization methods across high-concentration regimes have been proposed to shorten the data acquisition time. In this work, a data-driven encoder-decoder convolutional neural network (deep-ULM) and a model-based deep unfolded network embedding a sparsity prior (deep unfolded ULM) are analyzed in terms of localization accuracy and computational complexity. The results of simulated test data showed that both deep learning methods could handle overlapping PSFs better than centroid detection. Additionally, thanks to its model-based approach, deep unfolded ULM needed much fewer learning parameters and was computationally more efficient, and consequently achieved better generalizability than deep-ULM. It is expected that deep unfolded ULM will be more robust in-vivo.