NIH grant: Fullwave Physics-Informed Neural Networks for Rapid Ultrasound Simulation, Quantitative Imaging, and Adaptive Focusing (1R21EB039965-01)

retained text

NIH grant 1R21EB039965-01: Fullwave Physics-Informed Neural Networks for Rapid Ultrasound Simulation, Quantitative Imaging, and Adaptive Focusing
Institute: {'code': 'EB', 'abbreviation': 'NIBIB', 'name': 'National Institute of Biomedical Imaging and Bioengineering', 'admin_org_id': '12500035', 'admin_funding_url': ''}. PI: You  Li. Award: $186674.
Start: 2026-08-07T00:00:00.
Project Abstract
Ultrasound is one of the most widely used imaging modalities in medicine because it is safe, inexpensive, and
provides real-time visualization. However, current ultrasound remains largely qualitative and suffers from
reduced reliability when sound waves traverse heterogeneous tissues such as fat, muscle, or bone. In these
cases, reverberation, aberration, and nonlinear propagation degrade image quality, bias quantitative
measurements, and defocus transmit energy. Quantitative ultrasound (QUS) has the potential to generate
objective tissue property maps, such as speed-of-sound, attenuation, and density, that could improve diagnosis
(e.g., liver diseases) and guide therapies. Despite decades of research, QUS has not achieved broad clinical
adoption because most existing methods assume simplified physics, and existing full-wave solvers are too
computationally expensive for clinical use.
This project addresses these barriers by developing full-wave physics-informed recurrent neural networks
(RNN) that embed a differentiable solution of the nonlinear heterogeneous acoustic wave equation into a
recurrent neural network. In this framework, the recurrent cell implements the discretized full-wave update,
while the network weights correspond directly to tissue property maps. This single, interpretable model enables
three tightly linked capabilities: (i) rapid forward simulation of nonlinear propagation for artifact-aware sequence
design; (ii) quantitative inversion to recover speed-of-sound and attenuation maps under realistic conditions;
and (iii) adaptive focusing to optimize transmit phase and apodization through heterogeneous layers.
Aim 1 will develop a stable recurrent model of the nonlinear full-wave equation with multi-relaxation attenuation
and perfectly matched layers, validated against a trusted solver with <5% error and >10× GPU speedup.
Aim 2 will implement gradient-based quantitative inversion and adaptive focusing in silico, with target outcomes
of ≤10% relative mean absolute error in tissue acoustic property map estimation and ≥3 dB focal gain with
≥20% off-target energy reduction.
Aim 3 will experimentally validate and characterize the performance of the framework developed in Aims 1 and
2 using a research ultrasound scanner and tissue-mimicking acoustic phantoms and ex vivo tissue.
This project will deliver the first unified, physics-faithful, differentiable full-wave framework for ultrasound. By
bridging physics-based modeling with learning-based optimization, it establishes a new paradigm for
quantitative ultrasound that is fast, interpretable, and artifact-aware. Although validated initially in phantoms
and ex vivo tissues, the framework is organ-agnostic and has broad potential applications ranging from liver
disease assessment to transcranial imaging and image-guided therapy, directly aligned with NIBIB’s mission to
develop enabling biomedical imaging technologies.

document fetched 2026-08-16 06:34