Machine learning-enabled quantitative ultrasound techniques for tissue differentiation

Abstract

Purpose Quantitative ultrasound (QUS) infers properties about tissue microstructure from backscattered radio-frequency ultrasound data. This paper describes how to implement the most practical QUS parameters using an ultrasound research system for tissue differentiation. Methods This study first validated chicken liver and gizzard muscle as suitable acoustic phantoms for human brain and brain tumour tissues via measurement of the speed of sound and acoustic attenuation. A total of thirteen QUS parameters were estimated from twelve samples, each using data obtained with a transducer with a frequency of 5–11 MHz. Spectral parameters, i.e., effective scatterer diameter and acoustic concentration, were calculated from the backscattered power spectrum of the tissue, and echo envelope statistics were estimated by modelling the scattering inside the tissue as a homodyned K-distribution, yielding the scatterer clustering parameter α and the structure parameter κ. Standard deviation and higher-order moments were calculated from the echogenicity value assigned in conventional B-mode images. Results The k-nearest neighbours algorithm was used to combine those parameters, which achieved 94.5% accuracy and 0.933 F1-score. Conclusion We were able to generate classification parametric images in near-real-time speed as a potential diagnostic tool in the operating room for the possible use for human brain tissue characterisation.

Development of continuous assessment of muscle quality and frailty in older subjects using multi-parametric omics based on combined ultrasound and blood biomarkers: a study protocol for a cluster randomised controlled trial

“…Ultrasound quantitative biomarkers based on raw data have been previously generally shown to capture both tissue composition and microstructural properties, and have shown to encode a richer information content than ultrasound B-mode images in arti cial intelligence models [52,54,81,82]. Ultrasound spectroscopy parameters and tissue acoustic properties such as speed-of-sound and attenuation have been linked to tissue composition [47] and viscoelastic changes in muscle [49] in elderly subjects with sarcopenia. Particularly, speed-of-sound showed correlations with MRI adiposity estimates in the calf muscles [89], CT assessment of the psoas muscle [90] and short-term changes in muscle due to immobilization [91].…”

“…Ultrasound morphometric measurements of sarcopenia in older adults have shown mild to moderate associations with frailty [41]. More recently, several quantitative ultrasound techniques have emerged based on the analysis of echogenicity, texture parameters, elastography and acoustic wave properties, with still limited translation to clinical practice [42][43][44][45][46][47][48][49][50][51]. Arti cial intelligence is bringing new opportunities to objectivize musculoskeletal ultrasound, with recent works demonstrating automatic muscle segmentation and ber angle detection and textural discrimination of muscle microstructures [52][53][54][55].…”

Development of Continuous Assessment of Muscle Quality and Frailty in Older Patients Using Multiparametric Combinations of Ultrasound and Blood Biomarkers: Protocol for the ECOFRAIL Study

“…Attenuation coefficient [74,75] Measurement of tissue reflectivity after attenuation compensation. Backscattering coefficient [49] Spectroscopy measurement of backscattered signal variation with frequency, including parametrizations such as spectral slope, spectral intercept, and midband fit Power spectrum (Lizzi-Feleppa parameters) [49,76,77] Fitting of raw envelope signal to speckle statistical distribution models, including Rayleigh, Nakagami, and homodyned K-distribution. Estimation of scatterer concentration, spacing, and coherence from fitted model parameters.…”

“…More recently, various quantitative ultrasound techniques have surfaced, involving the analysis of echogenicity, texture parameters, elastography, and acoustic wave properties. However, their translation to clinical practice is still limited [42][43][44][45][46][47][48][49][50][51]. Artificial intelligence is presenting new opportunities to objectify musculoskeletal ultrasound, with recent studies showcasing automatic muscle segmentation, fiber angle detection, and textural discrimination of muscle microstructures [52][53][54][55].…”

HaTU-Net: Harmonic Attention Network for Automated Ovarian Ultrasound Quantification in Assisted Pregnancy

“…Figure 1 shows three examples of US images of ovary and follicles. Quantifying ultrasound images ensures reproducibility and reliability [ 6 , 7 ]. Ultrasound imaging artifacts impede the performance of deep learning-based segmentation methods.…”