Quantitative Ultrasound (QUS)

Updated 8 June 2026

Quantitative Ultrasound (QUS) is a framework that analyzes raw radiofrequency echoes using mathematical, statistical, and physics-based models to estimate key tissue parameters.
It employs techniques such as Nakagami and Homodyned K-distributions along with spectral models to generate quantitative parametric maps of scatterer properties, attenuation, and effective scatterer diameter.
Modern QUS methods integrate machine learning and physics-informed models to improve system calibration, overcome data heterogeneity, and enable real-time clinical applications.

Quantitative Ultrasound (QUS) is a framework for inferring tissue microstructure using mathematical analysis of the radiofrequency (RF) echo signals backscattered by biological structures in clinical ultrasound. Unlike conventional B-mode imaging, which is largely qualitative, QUS methods produce parametric maps of tissue that quantify features such as scatterer properties, acoustic attenuation, and the organization of microstructural elements. These quantitative biomarkers are derived by modeling and inverting the physical mechanisms that underpin echo formation, utilizing statistical, spectral, or physics-based parameterizations. QUS continues to advance as both a rigorous research area and an increasingly relevant set of clinical tools for noninvasive diagnosis, grading, and therapy monitoring.

1. Theoretical Foundations and Parameter Models

The core QUS objective is to estimate intrinsic tissue parameters—such as scatterer number density, effective scatterer diameter (ESD), backscatter coefficient (BSC), attenuation coefficient (AC), and first-order envelope statistics—from the raw RF signals. These estimated quantities are designed to be as independent as possible from imaging system settings and operator variability.

Statistical Envelope Models

Nakagami Distribution: Models the envelope amplitude using parameters $m$ (shape) and $\Omega$ (scale), with $m$ related to scatterer density and spatial organization. The probability density function (PDF) is:

$f(A;m,\Omega) = \frac{2 m^m}{\Gamma(m)\,\Omega^m}\;A^{2m-1}\;\exp\!\left(-\frac{m}{\Omega}A^2\right)$

where $A$ is the envelope amplitude and $\Gamma$ is the Gamma function (Tehrani et al., 2022, Poul et al., 2024).

Homodyned K-Distribution (HKD): Captures the statistics of speckle with both coherent and diffuse scattering components, parameterized by the scatterer clustering $\alpha$ and the coherent-to-diffuse power ratio $k = \epsilon^2/(2\sigma^2\alpha)$ (Tehrani et al., 2022). The PDF can be written as:

$p_{HK}(A|\epsilon,\sigma^2,\alpha) = A \int_0^\infty u J_0(u\epsilon) J_0(uA)\left(1+\frac{u^2\sigma^2}{2}\right)^{-\alpha} du$

where $J_0$ is the zeroth-order Bessel function (Tehrani et al., 2022).

Burr Distribution: A two-parameter ( $\Omega$ 0, $\Omega$ 1) model describing first-order speckle statistics, especially suitable when the scatterer size/density has a power-law distribution (Poul et al., 2024).

Frequency-Domain Spectral Models

Backscatter Coefficient (BSC): Expresses frequency-dependent scattering power, typically modeled:

$\Omega$ 2

where $\Omega$ 3 is scatterer density, $\Omega$ 4 is the scatterer size distribution, and $\Omega$ 5 is the form factor (e.g., Gaussian, Faran) (Jafarpisheh et al., 2021).

Attenuation Coefficient (AC): Expresses frequency-dependent amplitude loss, commonly postulated as $\Omega$ 6, where $\Omega$ 7 is the slope parameter (linear term) in dB/MHz·cm (Poul et al., 23 Mar 2026, Song et al., 2021).
Effective Scatterer Diameter (ESD): Derived by fitting the measured BSC against a theoretical form-factor model, typically assuming a Gaussian shape for scatterers (Nizam et al., 2019, Jafarpisheh et al., 2021).

These statistical and spectral QUS parameters serve as surrogates for tissue microstructure and are mapped in space to form parametric images.

2. Methods of QUS Parameter Estimation

A wide range of computational methods underpin QUS parameter estimation, including moment-based estimators, maximum-likelihood optimization, constrained inverse problems, machine learning, and hybrid deep learning–physics-informed algorithms.

Classic Statistical and Spectral Estimation

Patch-Based Methods: Classical QUS divides the data into overlapping 2D windows (“patches”) and assumes each patch is locally homogeneous (Tehrani et al., 2022, Tehrani et al., 2022). Parameters such as Nakagami $\Omega$ 8 and HKD $\Omega$ 9 are calculated either via maximum-likelihood (ML), method of moments, or specialized log-moment approaches (e.g., Destrempes’s XU estimator for HKD).
Spectral Difference and Reference Phantom Normalization: For attenuation and BSC estimation, sample spectra are normalized by phantom spectra to remove system transfer function influences and correct for attenuation (Poul et al., 2024, Deeba et al., 2021, Poul et al., 23 Mar 2026). AC is then estimated by fitting the depth-dependence of the log-spectrum slope, and BSC/ESD by regression against frequency.
Constrained Optimization: Modern QUS inverts linear or nonlinear relationships (e.g., system of equations for BSC vs. scatterer size distribution) using constrained least squares or quadratic programming, enforcing physical constraints such as non-negativity and sum-to-one for probability distributions (Jafarpisheh et al., 2021, Tehrani et al., 7 Jan 2025).

Machine Learning and Deep Learning

CNN and Segmentation Networks: Patchless pixel-level estimation of QUS maps, such as scatterer number density, is now feasible with deep convolutional networks trained on large-scale, system-diverse simulated datasets. Notable examples include Pyramid Attention Networks and DeepLabV3 (Tehrani et al., 2022, Tehrani et al., 2023, Tehrani et al., 2022).
Multi-Task Learning: Joint training on supervised main tasks (e.g., density segmentation) and auxiliary subtasks (e.g., Nakagami $m$ 0 regression) improves invariance to system settings and tissue intensity (Tehrani et al., 2022).
Transfer Function Approaches: Machine-to-machine spectral (frequency-domain) transfer functions, including Wiener-filter–inspired regularization, are used to harmonize data distributions across different scanners, enabling transferability of trained deep models (Soylu et al., 2023).
Physics-Informed Neural Networks and Deep Unfolding: Hybrid architectures embed full or partial physical forward models (e.g., finite-difference time-domain [FDTD] wave propagation) and learn only the update rules or regularization, combining interpretability with data efficiency and computational speed (Zhang et al., 4 Mar 2026).
Uncertainty Quantification: Bayesian neural networks yield not only point estimates of QUS parameters (e.g., HKD $m$ 1, $m$ 2) but also credible intervals, enabling reliable interpretation and adaptive data collection (Tehrani et al., 2022, Tehrani et al., 2023).

3. Parametric Imaging and Volumetric QUS

Quantitative maps of tissue parameters are formed by reconstructing the spatial distribution of QUS properties.

2D/3D Mapping: Early QUS produced 2D parametric maps via sliding-window estimates. Volumetric QUS uses spatially adaptive, weighted total variation (TV) regularization to jointly reconstruct attenuation, backscatter, and ESD in 3D, respecting both local homogeneity and tissue interfaces (Deeba et al., 2021).
Adaptive Regularization: The regularization weights in the TV norm can be modulated based on local envelope SNR deviation from Rayleigh speckle, allowing strong denoising within homogeneous regions and edge-preservation at heterogeneity boundaries (Deeba et al., 2021).
Multiparametric and Multimodal Fusion: Simultaneous mapping of ACE, IBC, and ESD provides superior tissue classification—e.g., perfect discrimination between normal and steatotic liver in in vivo studies—relative to any single parameter (Deeba et al., 2021).
Machine Learning Guided Patchless Mapping: Deep models can directly estimate per-pixel scatterer number density, speckle model parameters, and even deliver frame-wise or pixel-wise uncertainty (Tehrani et al., 2023, Tehrani et al., 2022).

4. Applications Across Anatomical and Clinical Contexts

QUS biomarkers have demonstrable utility in tissue characterization, cancer diagnostics, soft-tissue disease assessment, and monitoring of therapy response.

Tissue and Disease Characterization

Periodontal Soft Tissue Differentiation: Oral tissues such as gingiva versus alveolar mucosa can be sharply discriminated using Burr and Nakagami model parameters (with >90% accuracy and $m$ 3), correlating with histologically determined collagen content (Poul et al., 2024, Poul et al., 2024).
Tumor and Lymph Node Assessment: Envelope statistics derived from the HKD (e.g., $m$ 4 and $m$ 5) differentiate metastatic from non-metastatic lymph nodes and detect microstructural changes in response to radiation therapy in canine tumor models (Gardner et al., 25 Mar 2025).
Nerve Fascicle Microstructure: QUS-derived backscatter coefficient and envelope entropy show strong negative correlations with combined collagen and myelin content in human ulnar nerve fascicles, exceeding the sensitivity of Nakagami $m$ 6 or AC in high-frequency regimes (Byra et al., 2019).

Bone Quality and Musculoskeletal

Bone Quality Assessment in Scoliosis: The Frequency Amplitude Index (FAI), extracted from spectral peaks of the reflected signal, strongly correlates (R² = 0.824) with reflection coefficient, serving as an indicator for bone quality and curve progression risk in adolescent idiopathic scoliosis (Song et al., 2021).
Skeletal Status and Curve Progression: Lower FAI is associated with a higher risk of scoliosis curve progression in non-mild cases (30% vs 5%) (Song et al., 2021).

Liver and Oncologic Imaging

Liver Fibrosis/Steatosis: Volumetric, multiparametric QUS (3D mapping of ACE, IBC, ESD) correlates strongly with MRI-PDFF and perfectly discriminates steatotic from normal livers using simple classifiers on QUS features (Deeba et al., 2021).
Breast Cancer Classification: ESD and MSS, estimated via EEMD-based decomposition and AR spectral analysis, provide high diagnostic accuracy (AUC=0.96) for breast lesion classification (Nizam et al., 2019).

5. System Dependence, Reference Phantoms, and Calibration

A fundamental QUS challenge is the removal or correction of system dependence to ensure transferability, reproducibility, and clinical validity.

Reference Phantom Normalization: Essential for BSC, AC, and ESD estimation, as system transfer functions are normalized out by subtraction and division of log-spectra (Deeba et al., 2021, Jafarpisheh et al., 2021).
System Dependency in Statistical Parameters: Envelope-derived features (e.g., SNR, skewness, kurtosis) and speckle statistics are affected by imaging settings and window size; imaging of “reference” phantoms with known properties is used to calibrate and mitigate these dependencies (Tehrani et al., 2022).
Domain Adaptation in Deep Models: Adaptive Batch Normalization (AdaBN) reestimates only the mean/variance statistics of normalization layers using a small number of new-domain “reference” images, enabling rapid calibration to novel scanners (Tehrani et al., 2022).
Machine-to-Machine (M2M) Calibration: Deep learning models can be harmonized across scanners using transfer functions derived from calibration phantoms and Wiener-style spectral regularization, achieving cross-machine classification accuracies >88% (AUC≈0.99) (Soylu et al., 2023).

6. Challenges, Limitations, and Future Directions

QUS continues to face critical methodological, clinical, and computational challenges:

Data Scarcity and Heterogeneity: Small cohort sizes, inter-scanner variability, and lack of ground truth limit model generalizability and performance benchmarking (Zhou et al., 2020, Tehrani et al., 2022).
Resolution-Precision Trade-Off: Sliding window and patch-based approaches fundamentally trade spatial resolution for statistical confidence; adaptive and volumetric regularization, as well as patchless deep models, mitigate but do not eliminate this constraint (Deeba et al., 2021, Tehrani et al., 2023).
In Vivo and Cross-Domain Robustness: Simulated datasets may not fully capture in vivo complexity; explicit validation in realistic biological settings, including correlation to histopathology, remains essential (Tehrani et al., 2022, Tehrani et al., 2024).
Physics-Informed Learning: Future methods increasingly embed physics (acoustic wave propagation, scattering) within machine learning architectures, as exemplified by Deep Unfolded Full Waveform Inversion for speed-of-sound QUS (Zhang et al., 4 Mar 2026).
Computation and Real-Time Implementation: Advanced methods (e.g., DUFWI) show order-of-magnitude speed-ups over classical PDE inversion but require further optimization for real-time clinical deployment (Zhang et al., 4 Mar 2026).
Explainability and Uncertainty: Bayesian and uncertainty-aware models are emerging to quantify estimator confidence, flagging regions demanding more data and improving decision support (Tehrani et al., 2022, Tehrani et al., 2023).

QUS is evolving into a unified, multiparametric imaging framework integrating robust physics-based models, system calibration, deep learning, and careful validation, supporting both the next generation of research and translation to the clinic.