KAN-ULM: Advancing Super Resolution Imaging in Ultrasound Localization Microscopy Through Compact Deep Learning Model

Ultrasound Localization Microscopy (ULM) has gained recognition as an advanced imaging technique capable of visualizing microvasculature with exceptional detail, offering critical insights into cerebral blood flow dynamics. The ULM pipeline involves multiple computationally intensive stages, which significantly slow down the process. The localization of microbubbles (MBs) is a critical step that enhances the accuracy of MB tracking. In this study, we explore Kolmogorov-Arnold Networks (KAN) and present KAN-ULM, a highly compact deep network that optimizes the localization step in ULM. Our research systematically analyzes various configurations of KAN against a well-defined metric, providing valuable insights for optimizing network architecture. Despite operating within a limited parameter range, our results demonstrate that KAN achieves remarkable resolution, surpassing other state-of-the-art methods, showcasing its potential in high-resolution imaging applications.Clinical relevance- The KAN-ULM model significantly enhances MB localization in ULM, which could potentially enable finer visualization of the microvasculature and support more accurate diagnoses and personalized treatments in future clinical applications.