Fast Large-Scale, Non-Rigid Motion Estimation for ULM

Motion correction of ultrasound (US) sequences is essential to avoid resolution deterioration in Ultrasound Localization Microscopy (ULM) because ULM sequences are usually long and can contain considerable motion, particularly in clinical applications at non-rigid tissue and when transducers are hand-held. However, conventional approaches for motion estimation are increasingly time-consuming when large motions are present. To quickly estimate large-scale, non-rigid motion but with sub-pixel accuracy, we tested and modified the neural network Recurrent All-Pairs Field Transforms (RAFT). For supervised learning, we implemented a simulation pipeline to generate realistic motion-corrupted US sequences. RAFT not only achieved sub-pixel estimation accuracy comparable to conventional approaches, but also was considerably faster and adapted well to in-vivo data although trained on synthetic data.