Deep 3D convolutional neural networks for fast super-resolution ultrasound imaging

Super-resolution ultrasound imaging (SR-US) is a new technique which breaks the diffraction limit and can help visualize microvascularity at a resolution of tens of microns. However, image processing methods for spatiotemporal filtering needed in SR-US for microvascular delineation, such as singular value decomposition (SVD), are computationally burdensome and must be performed off-line. The goal of this study was to evaluate a novel and fast method for spatiotemporal filtering to segment the microbubble (MB) contrast agent from the tissue signal with a trained 3D convolutional neural network (3DCNN). In vitro data was collected using a programmable ultrasound (US) imaging system (Vantage 256, Verasonics Inc, Kirkland, WA) equipped with an L11-4v linear array transducer and obtained from a tissue-mimicking vascular flow phantom at flow rates representative of microvascular conditions. SVD was used to detect MBs and label the data for training. Network performance was validated with a leave-one-out approach. The 3DCNN demonstrated a 22% higher sensitivity in MB detection than SVD on in vitro data. Further, in vivo 3DCNN results from a cancer-bearing murine model revealed a high level of detail in the SR-US image demonstrating the potential for transfer learning from a neural network trained with in vitro data. The preliminary performance of segmentation with the 3DCNN was encouraging for real-time SR-US imaging with computation time as low as 5 ms per frame.