Model-based Deep Learning on Ultrasound Channel Data for Fast Ultrasound Localization Microscopy

Ultrasound localization microscopy (ULM) can break the diffraction limit of ultrasound imaging. However, a long data acquisition time is often required due to the use of low concentrations of microbubbles (MBs) for high localization accuracy. Lately, deep learning-based methods that can robustly localize high concentrations of microbubbles (MBs) have been proposed to overcome this constraint. In particular, deep unfolded ULM has shown promising results with a few parameters by using a sparsity prior. In this work, deep unfolded ULM is further extended to perform beamforming as well as MB localization. The proposed network learns data-dependent beamforming weights that are optimal for deep unfolded ULM to locate MBs. The images beamformed by the network were sharper than delay-and-sum beamformed images. In a simulated test set at an MB density of 3.84 mm−1, the proposed network reconstructed 87 % of MBs with the precision of 0.99 while achieving comparable localization accuracy to deep unfolded ULM, when centroid detection and deep unfolded ULM reconstructed 42 % and 67 % of MBs with the precision of 0.75 and 0.99, respectively.