A General Deep Learning Model for Ultrasound Localization Microscopy

Abstract

With the help of microbubbles (MBs), ultrasound localization microscopy (ULM) breaks the diffraction limit of ultrasound imaging. Accurate and robust localization of MBs is central to ULM but still challenged by ultrasound point spread function (PSF) analysis and potential overlapping and clustering of MBs even at a low MB concentration. Current deep learning (DL) methods enable fast acquisitions and high localization rates via prior knowledge of PSFs, consistency between training and test data, or image pre-processing, thus hampering model generality. We hereby propose a general system-independent DL-based ULM framework, which encompasses (1) generation of synthetic ultrasound MB images with a variety of PSFs and medium complexity as training data and (2) two cascaded DL models for de-speckling and deconvolution, respectively, to achieve MB localization. The proposed framework not only achieved a precision of 0.74±0.06 and a recall of 0.68±0.08 at the threshold of half wavelength, but also distinguished overlapped MBs on test data with unseen distinct PSFs provided by 2022 IUS Ultra-SR challenge.