Super-Resolution Ultrasound Image Filtering with Machine-Learning to Reduce the Localization Error

Localization-based super-resolution imaging requires accurate detection of spatially isolated microbubbles. The reason for this requirement is that interfering or overlapping signals resulting from multiple microbubbles within the resolution limit can cause position errors. In addition to this, noise and artefacts (e.g. residual tissue signal after tissue-microbubble separation) further reduce the quality and hence the spatial resolution in SR imaging. Therefore, correctly identifying the echoes as noise, single microbubble, multiple microbubbles, or artefact is important. In this study, the use of fast classification methods for identification and rejection of non-single microbubble echoes were demonstrated. Most commonly used supervised classification methods, including Decision Trees, Discriminant Analysis, Logistic Regression, Support Vector Machine, Ensembles, k-Nearest Neighbors, and Naive Bayes, were implemented for filtering artefacts and noise in super-resolution ultrasound images. Results showed that the Ensemble method, explicitly designed to deal with unbalanced data, achieved the best result since most of the localized events are true microbubbles, which is typical for super-resolution imaging datasets.