Deep Learning-Based Super-resolution Ultrasound Speckle Tracking Velocimetry

Abstract

Deep ultrasound localization microscopy (deep-ULM) allows sub-wavelength resolution imaging with deep learning. However, the injection of contrast agents (CAs) in deep-ULM is debatable because of their potential risk. In this study, we propose a deep learning-based super-resolution ultrasound (DL-SRU), which employs the concept of deep-ULM and a convolutional neural network. The network is trained with synthetic tracer images to localize positions of red blood cells (RBCs) and reconstruct vessel geometry at high resolution, even for CA-free ultrasound (US) images. The proposed algorithm is validated by comparing the full width at half-maximum values of the vascular profiles reconstructed by other techniques, such as the standard ULM and the US average intensity under in silico and in vitro conditions. RBC localization by DL-SRU is also compared with that by other localization approaches to validate its performance under in vivo condition, especially for veins in the human lower extremity. Furthermore, a two-frame particle tracking velocimetry (PTV) algorithm is applied to DL-SRU localization for accurate flow velocity measurement. The velocity profile obtained by applying the PTV is compared with a theoretical value under in vitro condition to verify its compatibility with the flow measurement modality. The velocity vectors of individual RBCs are obtained to determine the applicability to in vivo conditions. DL-SRU can achieve high-resolution vessel morphology and flow dynamics in vasculature, mapping 110 super-resolved images per second on a standard PC, regardless of various imaging conditions. As a result, the DL-SRU technique is much more robust in localization compared with previous deep-ULM. In addition, the performance of DL-SRU is nearly the same as that of deep-ULM in rapid computational processing and high measurement accuracy. Thus, DL-SRU might become an effective and useful instrument in clinical practice.