Modified Residual Dense Network based super-resolution localization method for high-concentration microbubbles

The practical limitation of ultrasound localization microscopy for clinical translation is the trade-off between microbubble concentration and data acquisition time. Recently, deep learning-based approaches have shown promising capability in microbubble localization accuracy when using a high-concentration microbubble injection to shorten acquisition time. In this study, we construct a Modified Residual Dense Network (MRDN) for high-concentration microbubble super-resolution localization. By subjecting the collected data to non-local mean filtering operations, the MRDN is used for continuous learning. This method can be well used at high concentrations of 16 mathbf{mm}{ boldsymbol{-2}} with a high localization accuracy (localization error boldsymbol{:22.3 mu} mathbf{m}) and high localization reliability (Jaccard index:0.78).