Ultrasound Multiple Point Target Detection and Localization using Deep Learning

Abstract

Super-resolution imaging (SRI) can achieve subwavelength resolution by detecting and tracking intravenously injected microbubbles (MBs) over time. However, current SRI is limited by long data acquisition times since the MB detection still relies on diffraction-limited conventional ultrasound images. This limits the number of detectable MBs in a fixed time duration. In this work, we propose a deep learning-based method for detecting and localizing high-density multiple point targets from radio frequency (RF) channel data. A Convolutional Neural Network (CNN) was trained to return confidence maps given RF channel data, and the positions of point targets were estimated from the confidence maps. RF channel data for training and evaluation were simulated in Field II by placing point targets randomly in the region of interest and transmitting three steered plane waves. The trained CNN achieved a precision and recall of 0.999 and 0.960 on a simulated test dataset. The localization errors after excluding outliers were within ± 46 µm and ± 27 µm in the lateral and axial directions. A scatterer phantom was 3-D printed and imaged by the Synthetic Aperture Real-time Ultrasound System (SARUS). On measured data, a precision and recall of 0.976 and 0.998 were achieved, and the localization errors after excluding outliers were within ± 101 µm and ± 75 µm in the lateral and axial directions. We expect that this method can be extended to highly concentrated microbubble (MB) detection in order to accelerate SRI.

3D printed calibration micro-phantoms for super-resolution ultrasound imaging validation

“…Phantoms could be developed with scatterers placed much closer, to tune algorithms to be able to separate signals from partially overlapping reflections. This would be highly relevant for instance for some of the new types of SRI schemes which seek to be able to separate reflections much closer than the wavelength [32].…”

“…The phantom feature sizes are by no means the limit of the printing system. In [32] we presented another 3D printed phantom containing 45 × 45 × 1000 µm 3 scatterers for 2D imaging. Integration of the signal across the elevation plane allows for an increased intensity even though the scatterers were significantly smaller in cross-section than those used in this work.…”

Deep Learning Models for Fast Ultrasound Localization Microscopy

“…To obtain the MB positions in the higher-resolution grid, the first up block upsamples the features by a factor of 2 and the other up blocks perform upsampling by a factor of 4. A detailed description of down, conv, and up blocks can be found in [15], [16].…”

A Deep Learning Framework for Spatiotemporal Ultrasound Localization Microscopy

“…Others investigated the application of deep learning-based algorithms to enhance the localization of individual MB when higher concentrations are used. Those were either based on radiofrequency (RF) data [13] or envelope-detected images [14], [15], [16], [17] and all relied on a per-frame localization.…”