Deep Learning-Based Microbubble Localization for Ultrasound Localization Microscopy

Abstract

Ultrasound localization microscopy (ULM) is an emerging vascular imaging technique that overcomes the resolution-penetration compromise of ultrasound imaging. Accurate and robust microbubble (MB) localization is essential for successful ULM. In this study, we present a deep learning (DL)based localization technique that uses both Field-II simulation and in vivo chicken embryo chorioallantoic membrane (CAM) data for training. Both radiofrequency (RF) and in-phase quadrature (IQ) data were tested in this study. The simulation experiment shows that the proposed DL-based localization was able to reduce both missing MB localization rate and MB localization error. In general, RF data showed better performance than IQ. For the in vivo CAM study with high MB concentration, DL-based localization was able to reduce the vessel MB saturation time by more than 50% as compared to conventional localization. Additionally, we propose a DL-based framework for real-time visualization of the high-resolution microvasculature. The findings of the paper support the use of DL for more robust and faster MB localization, especially under high MB concentrations. The results indicate that further improvement could be achieved by incorporating temporal information of the MB data.

Deep Learning-Based Image Analysis for the Quantification of Tumor-Induced Angiogenesis in the 3D In Vivo Tumor Model—Establishment and Addition to Laser Speckle Contrast Imaging (LSCI)

“…Unlike our presented work, it does not include biological factors such as tumor-induced angiogenesis or the effects of anti-angiogenic substances. Other recent and more technical publications include the use of the CAM model for the application of ultrasound localization microscopy, which utilizes a CNN to visualize microvessels [ 59 , 60 ]. Li et al talk about the difficulty of obtaining the large in vivo training datasets required to train previously used cross-correlation-based localization methods combatted by the improved learning scheme of CNNs, though only using ex ovo CAM assay training data.…”

Three-Dimensional Ultrasound Localization Microscopy with Bipartite Graph-Based Microbubble Pairing and Kalman-Filtering-Based Tracking on a 256-Channel Verasonics Ultrasound System with a 32 × 32 Matrix Array

“…injection of high MB concentration or in-paint the ULM image); they suffer from the uncertainty of the number of iterations, inaccurate estimated point spread function (PSF) and exhaustive parameter optimization to achieve an optimal image quality. Additionally, 2D deep learning based ULM (deep-ULM) approaches [18][19][20][21][22][23][24] have been actively developed recently to improve this issue. Deep-ULMs used neural networks to extract features of the MB signal during the training process and then make predictions to identify MB signals, resulting in fast recognition of MB signals and microvascular reconstruction even at high MB concentrations.…”

Super resolution ultrasound imaging Using deep learning based micro-bubbles localization

“…1) Training data generation: First we would like to point out unlike methods shown in [12], training data generated by simulations were not employed in our experiments for following consideration. It seems that oscillations of MBs are not considered in above mentioned paper and essentially, MBs and normal scatters suffer different underlying physical disciplines when insonified by ultrasound beam [13].…”

“…3) Loss function: Besides conventional L 1 loss [12], dice loss was added to the total loss, increasing the convergence speed of training in our experiments. When using dice loss, one threshold was pre-set to segment the heatmap (label) to foreground and background.…”