Real-Time Model-Based Quantitative Ultrasound and Radar

TL;DR

This work tackles the slow, locally-minimizing nature of full-waveform inversion (FWI) for quantitative ultrasound and radar imaging by proposing MB-QRUS, a model-based neural network that unfolds the FWI gradient using a physics-informed U‑Net. It introduces a time-domain, tensor representation of channel data and learns a gradient tensor to update multiple physical-property maps from data acquired with only eight elements, enabling real-time reconstruction for diverse transmission setups including linear probes. Across radar and US experiments on brain slices, MNIST-shaped phantoms, random ovals, and real liver data, MB-QRUS significantly outperforms FWI in accuracy (NRMSE, PSNR, SSIM) and speed (sub-second inference versus minutes to hours for FWI). The results demonstrate the method’s potential for fast, multi-parameter quantitative imaging in clinical contexts such as fast stroke imaging and fatty liver assessment, while maintaining versatility to handle surrounding-object and non-time-harmonic transmissions.

Abstract

Ultrasound and radar signals are highly beneficial for medical imaging as they are non-invasive and non-ionizing. Traditional imaging techniques have limitations in terms of contrast and physical interpretation. Quantitative medical imaging can display various physical properties such as speed of sound, density, conductivity, and relative permittivity. This makes it useful for a wider range of applications, including improving cancer detection, diagnosing fatty liver, and fast stroke imaging. However, current quantitative imaging techniques that estimate physical properties from received signals, such as Full Waveform Inversion, are time-consuming and tend to converge to local minima, making them unsuitable for medical imaging. To address these challenges, we propose a neural network based on the physical model of wave propagation, which defines the relationship between the received signals and physical properties. Our network can reconstruct multiple physical properties in less than one second for complex and realistic scenarios, using data from only eight elements. We demonstrate the effectiveness of our approach for both radar and ultrasound signals.

Figures (13)

• Figure 1: Illustration of the grid setup and data creation. (a)-(c) show an example of the grid setup of a brain with a stroke and 8 antennas surrounding the brain, while (d)-(f) show an example of a US setup with a linear probe and two circles with different physical properties. (b)-(c) demonstrate the wave propagation from one antenna over the grid for two successive time samples, similarly to (e)-(f) for US. (g) displays the creation process of the data, when the CD is created from the simulated physical properties, a known pulse (that defines $\mathbf{S}(t,x,z)$), and utilizing the wave propagation \ref{['eq:us Wave discrete']} and, \ref{['eq:radar wave disc']}.
• Figure 2: Inference time. The goal of the network is to reconstruct the physical properties mapping from the CD signals.
• Figure 3: An Example of MB-QRUS architecture for US case. The inputs to the network are initial guesses for the properties and the measured CD. The input to the U-Net block is the CD differences and the output is the gradient tensor $\mathbf{G}$. The output channels number after each convolution is presented.
• Figure 4: Radar properties reconstruction, by MB-QRUS and FWI compared to the GT for 4 test cases of a realistic brain slice with different orientations and a random stroke.
• Figure 5: Radar properties reconstruction by MB-QRUS and FWI compared to the GT for 4 cases of a scatter object using MNIST digits shapes (0, 1, 6, and 5), using fixed initialization method.