Ultrasound Microbubbles Localization Using Object Detection Model

Ultrasound Localization Microscopy (ULM) is a newly developing imaging technique capable of achieving sub-wave diffraction resolution for micro-vascular structures. However, traditional single microbubble (MB) localization methods have problems with precision, robustness, and computational efficiency. In this work, we propose a convolutional neural network (CNN) to localize MBs by converting the localization problem to an object detection problem. To identify the location of MBs in an image, our proposed network progressively extracted the features and output the coordinates of the detected MBs. This work compared the proposed method and Gaussian fit model on ultrasound images with realistic flow characteristics, generated by the IUS ULTRASR Challenge, using Jaccard index and localization precision as metrics. The proposed method outperformed the Gaussian fit model in terms of Jaccard index, precision, and processing time.