Traditional acoustic imaging algorithms rely on a clean, elegant geometric assumption: distance equals time multiplied by velocity ($d = v \cdot t$). To map an echo received by a transducer array back to its precise origin in space, standard beamforming techniques—such as Delay-and-Sum (DAS)—assume that the speed of sound ($v$) is uniform across the entire medium. In human soft tissue, this is typically approximated as a constant 1540 m/s.
However, real biological tissue is highly heterogeneous. Fat, muscle, glandular tissue, and lesions all possess distinct densities and acoustic impedances, as noted by Chi‐Fai Ng et al. in their study on energy wave propagation (The effect of renal cortical thickness on the treatment outcomes of kidney stones treated with shockwave lithotripsy). As an acoustic wavefront travels through these varying layers, its local velocity fluctuates, causing the wave to refract, bend, and desynchronize. When a reconstruction algorithm forces a constant velocity assumption onto a variable medium, it miscalculates the time-of-flight, resulting in wavefront distortion or acoustic aberration.
Assumed Homogeneous Path: [Transducer] -------- (Straight Ray) --------> [Target] (Calculated t = d / 1540)
Actual Heterogeneous Path: [Transducer] --(Fat: Slow)--\__(Muscle: Fast)--> [Target] (Actual t != Calculated t)
This mismatch degrades image quality across multiple modalities. It introduces phase and amplitude distortions that cause blurring, artifacts, and a severe loss of spatial resolution and contrast (Yu-chen Zengqiu et al. in Iterative Pulse–Echo Tomography for Ultrasonic Image Correction). In optoacoustic imaging, uncorrected propagation delays lead to severe misalignment during image reconstruction (Firat Ozdemir et al. in OADAT: Experimental and Synthetic Clinical Optoacoustic Data for Standardized Image Processing), while in super-resolution ultrasound localization microscopy, these aberrations compromise sub-wavelength tracking accuracy (Marcelo Lerendegui et al. in ULTRA-SR Challenge: Assessment of Ultrasound Localization and TRacking Algorithms for Super-Resolution Imaging).
To overcome these limitations, researchers are developing methods to estimate the true, spatially varying speed of sound (SoS) or to bypass the uniform velocity constraint entirely.
Tomographic and Statistical Estimation
One direct way to fix the wavefront is to map the variations in the medium's velocity explicitly. By treating the imaging problem tomographically, the system can calculate an updated velocity map and re-compute the true acoustic paths.
The authors of Iterative Pulse–Echo Tomography for Ultrasonic Image Correction address this using an iterative Computed Ultrasound Tomography in Echo Mode (CUTE) framework. Instead of relying on a single-pass estimation, which often struggles in highly complex tissue structures, their method progressively refines the local SoS estimates. By iteratively minimizing the focus errors caused by acoustic aberrations, the algorithm restores wavefront coherence and sharpens the final image.
In many clinical scenarios, however, mapping these boundaries explicitly is difficult due to the absence of bright, isolated acoustic reflections to act as calibration targets. To solve this, Laure Giovangigli et al. proposed a mathematical framework that leverages the inherent "speckle"—the grainy, noise-like interference patterns seen in ultrasound images (Probing the speckle to estimate the effective speed of sound, a first step towards quantitative ultrasound imaging). By treating the tissue as a random, multi-scale medium and applying stochastic homogenization, the authors demonstrated that the effective local speed of sound can be recovered purely from the statistical properties of the background speckle, removing the need for explicit structural targets.
Data-Driven and Deep Learning Beamforming
An alternative paradigm replaces explicit geometric ray-tracing with data-driven models. Rather than manually calculating complex wave refraction through varying tissue layers, deep neural networks can be trained to learn the underlying physics of wave propagation implicitly.
For explicit mapping, Walter Simson et al. trained convolutional neural networks on large, quasi-realistic simulated datasets (Investigating pulse-echo sound speed estimation in breast ultrasound with deep learning). Their network directly generates accurate, continuous sound speed maps from raw pulse-echo ultrasound data, successfully distinguishing between the distinct acoustic velocities of fat, glandular tissue, and lesions.
Taking this concept to its logical conclusion, Shujaat Khan et al. introduced "DeepBF," an end-to-end deep learning beamformer designed to replace the traditional DAS algorithm entirely (Deep Learning-based Universal Beamformer for Ultrasound Imaging).
Traditional Pipeline: [Raw RF Signals] ---> [DAS Beamformer (Assumes v=1540)] ---> [Blurred Image]
DeepBF Pipeline: [Raw RF Signals] ---> [Deep Neural Network (Learns v(x))] ---> [Sharp Image]
Instead of relying on rigid, predefined time-delay equations that break down in heterogeneous media, the single neural network directly transforms raw radiofrequency signals into high-quality images. The network implicitly learns to compensate for complex acoustic properties and tissue-dependent velocity variations, maintaining high resolution even under significant data subsampling rates.
Coherent Interferometry in Random Media
When the imaging environment becomes exceptionally chaotic—such as imaging through highly scattering media or turbulent backgrounds—deterministic tracking of sound speed variations becomes mathematically intractable. In these extreme cases, researchers rely on statistical interferometry to cancel out phase distortions.
In a single-author work, Liliana Borcea introduced High-resolution Coherent INTerferometry (HCINT) to image through media with random wave speed fluctuations (High-Resolution Interferometric Synthetic Aperture Imaging in Scattering Media). Rather than backpropagating waves based on absolute travel times, the method processes the cross-correlations of the data received at different sensors.
Liliana Borcea expanded on this in Imaging in Random Media by Two-Point Coherent Interferometry, demonstrating that analyzing the cross-correlations between pairs of points yields images that are "statistically stable." Because nearby acoustic paths experience nearly identical velocity distortions from the intervening medium, cross-correlating the signals mathematically cancels out the random phase aberrations. This allows the system to resolve sharp target images despite severe, unpredictable inhomogeneities in the propagation medium.
Key Foundational Work
Adaptive and robust techniques (ART) for thermoacoustic and photoacoustic tomography (2007) — contributed 1 concept elements to this direction.
Hybrid Medical Imaging: Scanning Thermoacoustic Tomography (2000) — contributed 1 concept elements to this direction.