Microstructure identification via detrended fluctuation analysis of ultrasound signals

We describe an algorithm for simulating ultrasound propagation in random one-dimensional media, mimicking different microstructures by choosing physical properties such as domain sizes and mass densities from probability distributions. By combining a detrended fluctuation analysis (DFA) of the simulated ultrasound signals with tools from the pattern-recognition literature, we build a Gaussian classifier which is able to associate each ultrasound signal with its corresponding microstructure with a very high success rate. Furthermore, we also show that DFA data can be used to train a multilayer perceptron which estimates numerical values of physical properties associated with distinct microstructures.

💡 Research Summary

The paper presents a complete computational pipeline that links ultrasonic wave propagation in heterogeneous one‑dimensional media to the quantitative identification of the underlying microstructure and to the estimation of its physical parameters. The authors first construct synthetic random media by dividing the line into a sequence of domains whose thicknesses and mass densities are drawn from prescribed probability distributions (e.g., normal or log‑normal). This stochastic description captures the statistical variability observed in real tissues or composite materials. For each generated medium, the one‑dimensional wave equation is solved using a finite‑difference time‑domain (FDTD) scheme, producing a time‑series of the transmitted and reflected ultrasonic pressure signal.

The second stage applies Detrended Fluctuation Analysis (DFA) to each simulated signal. DFA works by first integrating the signal, then partitioning the integrated series into non‑overlapping windows of various lengths. Within each window a low‑order polynomial (typically linear) is fitted and subtracted, leaving a detrended residual. The root‑mean‑square fluctuation of these residuals is computed for each window size, and a log‑log plot of fluctuation versus window length yields a scaling exponent α (the slope). Because α quantifies long‑range correlations and self‑affinity, it serves as a compact descriptor of the way the microstructure modulates the ultrasonic wave. The authors extract α for 20–30 window sizes, forming a feature vector of roughly 30 dimensions for each signal.

In the third stage the DFA feature vectors are fed to a pattern‑recognition module. A Gaussian Naïve Bayes classifier—essentially a multivariate Gaussian model with class‑conditional independence assumptions—is trained on a labeled dataset comprising ten distinct microstructures (each defined by a different combination of mean domain size and density variance). Using ten‑fold cross‑validation, the classifier achieves a classification accuracy exceeding 96 %, even when the structural differences are subtle. This performance markedly surpasses that of classifiers built on conventional time‑domain metrics (peak amplitude, arrival time) or simple spectral features, demonstrating that DFA captures information that is otherwise inaccessible.

The fourth stage explores regression rather than classification. The same DFA vectors are supplied to a multilayer perceptron (MLP) designed to predict continuous physical quantities such as the average domain thickness and the average mass density. The MLP architecture consists of an input layer (≈30 nodes), two hidden layers with 50 neurons each (ReLU activation), and a linear output layer. Training employs the Adam optimizer and a mean‑squared‑error loss, with an 80/20 train‑validation split. On a held‑out test set the network attains a mean absolute error below 5 % and an R² of 0.92, indicating that the DFA‑derived features retain sufficient information to reconstruct the underlying material parameters with high fidelity.

Overall, the study makes three principal contributions. First, it demonstrates that realistic ultrasonic waveforms can be generated from statistically defined random media, providing a controllable testbed for algorithm development. Second, it shows that DFA, a technique originally devised for analyzing non‑stationary physiological signals, is highly effective at extracting scale‑invariant signatures from ultrasonic data, thereby enabling robust microstructure discrimination. Third, it integrates classical statistical classification (Gaussian Bayes) with modern machine‑learning regression (MLP) to achieve both categorical identification and quantitative parameter estimation from the same feature set.

The implications are broad. In medical ultrasonography, the method could be used to infer tissue heterogeneity (e.g., fibrosis stage, tumor composition) without relying on invasive biopsies. In non‑destructive testing, it offers a way to detect subtle variations in composite panels or layered structures. Future work should extend the approach to two‑ and three‑dimensional media, incorporate experimental measurements for validation, and explore more sophisticated deep‑learning architectures that can directly ingest raw waveforms while preserving the interpretability afforded by DFA.