Speaker
Description
Super-resolution ultrasound imaging represents a significant advancement in the
field of medical imaging, particularly through the application of ultrasound localization
microscopy (ULM). However, ULM's requires sparse microbubble distributions, which in
turn, imposes prolonged acquisition times to adequately image the full
microvasculature. We introduce a novel super-resolution methodology that
overcomes traditional ULM limitations by employing a direct deconvolution technique
on all radiofrequency (RF) channels obtained by single plane-wave imaging. We
leverage a deep learning physics-based approach, which integrates Stolt's FK
migration beamforming algorithm into the network architecture. Focusing on
low-frequency ultrasound (1.7 MHz), our research targets deep imaging capabilities
(up to 10 cm) within a dense cloud of monodisperse microbubbles, accommodating
up to 1000 microbubbles within the 2D measurement volume. Our approach utilizes
a simulation framework that encompasses a broad spectrum of acoustic pressures
(5-250 kPa) to accurately model the complex, nonlinear response of resonant,
lipid-coated microbubbles. Previously, published work uses single-channel
deconvolution with a 1D dilated convolutional network. The core innovation of this
study lies in replacing this 1D network by various 2D architectures operating on all
RF lines simultaneously in order to exploit the correlation among transducer
elements and allowing end-to-end learning of microbubble location. We expect this
approach to reduce detection uncertainty and to more effectively handle high bubble
concentrations, leading to better localization in more realistic scenarios.
