Elsevier

Ultrasonics

Volume 114, July 2021, 106419
Ultrasonics

Classification of red blood cell aggregation using empirical wavelet transform analysis of ultrasonic radiofrequency echo signals

https://doi.org/10.1016/j.ultras.2021.106419Get rights and content

Highlights

  • We propose an adaptive analysis method of ultrasonics RF echo signals of blood.

  • RF signals are decomposed into a series of EMFs, from which DEMFs are selected.

  • Optimum parameters and combination of characteristics are searched for classifiers.

  • The machine learning technique is useful for the RBC aggregation classification.

Abstract

Grading red blood cell (RBC) aggregation is important for the early diagnosis and prevention of related diseases such as ischemic cardio-cerebrovascular disease, type II diabetes, deep vein thrombosis, and sickle cell disease. In this study, a machine learning technique based on an adaptive analysis of ultrasonic radiofrequency (RF) echo signals in blood is proposed, and its feasibility for classifying RBC aggregation is explored. Using an adaptive empirical wavelet transform (EWT) analysis, the ultrasonic RF signals are decomposed into a series of empirical mode functions (EMFs); then, dominant empirical mode functions (DEMFs) are selected from the series. Six statistical characteristics, including the mean, variance, median, kurtosis, root mean square (RMS), and skewness are calculated for the locally normalized DEMFs, aiming to form primary feature vectors. Random forest (RDF) and support vector machine (SVM) classifiers are trained with the given feature vectors to obtain prediction models for RBC classification. Ultrasonic RF echo signals are acquired from five groups of six types of porcine blood samples with average numbers of aggregated RBCs of 1.04, 1.20, 1.83, 2.31, 2.72, and 4.28, respectively, to test the classification performance of the proposed method. The best subset with regard to the variance, kurtosis, and RMS is determined according to the maximum accuracy based on the RDF and SVM classifiers. The classification accuracies are 84.03 ± 3.13% for the RDF classifier, and 85.88 ± 2.99% for the SVM classifier. The mean classification accuracy of the SVM classifier is 1.85% better than that of the RDF classifier. In conclusion, the machine learning method is useful for the discrimination of varying degrees of RBC aggregation, and has potential for use in characterizing and monitoring the RBC aggregation in vessels.

Introduction

Red blood cell (RBC) aggregation refers to separated RBCs aggregating into a “money string” structure, and is one of the main factors in the non-Newtonian fluidity of blood [1]. In general, RBC aggregation is a normal and reversible physiological phenomenon [1], [2]; however, excessive RBC aggregation influences microcirculation and the nutrient supply, and is common in many clinical diseases, such as ischemic cardio-cerebrovascular disease, type II diabetes, deep vein thrombosis, and sickle cell disease [3], [4], [5]. Thus, the ability to grade RBC aggregation quickly and noninvasively is of great clinical significance.

Owing to its noninvasive, real-time, and reduplicative advantages, an ultrasonic technique has been widely applied in investigations on the characterization and classification of RBC aggregation. Previously, RBC aggregations were assessed based on ultrasonic echogenicity [6], [7], [8]. In this method, the echogenic profiles of blood were estimated from ultrasonic images obtained using general commercial ultrasonic systems or specially designed measuring instruments. Subsequently, different parameters were estimated based on the ultrasonic backscattered signals and a backscattering coefficient (BSC) to characterize and quantify the RBC aggregation [9], [10], [11], [12], [13], [14], [15], [16]. Using an A-mode ultrasonic apparatus, the mean size of the RBC aggregates was estimated from the intensity of the backscattered signals of blood to assess the degree of RBC aggregation [9]. The spectral slope of the ultrasound backscattered signals was estimated to quantify the degree of RBC aggregation [10]. Cloutier et al. proposed a structure factor size estimator [11], which was then approximated using a structure factor model [12] with two indices (packing factor and aggregate diameter) to parameterize the ultrasonic BSC aggregation. Subsequently, the structure factor size was incorporated into an attenuation estimator cellular imaging method to parameterize the BSC with three parameters (packing factor, mean fractal aggregate diameter, and total attenuation) [13], [14]. To better approximate the structural characteristics of RBC aggregation, an effective medium theory combined with a structure factor model has been proposed for determining the aggregate size, aggregate compactness, and systemic hematocrit of RBCs [15], [16]. In addition, parameters from the Nakagami statistical model [17] and the probability distribution of the speckle intensity [18] have also been used to assess RBC aggregation. However, accurately grading RBC aggregation using analytical and numerical models remains challenging.

Ultrasonic tissue classification based on machine learning algorithms has been an active field of study in recent years. In these studies, features were extracted from decomposed ultrasonic radiofrequency (RF) echo signals reflecting the microstructures of tissues, and then tissue classification was performed with trained classifiers [19], [20], [21], [22], [23], [24]. Yoon et al. estimated the spectra of ultrasonic RF echo signals based on a fast Fourier transform algorithm, and then obtained the spectrum slope, midband fit, and intercept to assess the degree of RBC aggregation in a static condition [19]. Granchi et al. used a Morlet wavelet convoluted with ultrasonic RF echo signals to obtain 24 sub-band components with a frequency bandwidth of 1 MHz. Using this data, locally averaged hyperspace coefficients were estimated to classify blood samples of different concentrations [20], and two types of breast lesion tissues (invasive ductal carcinoma and fibroadenoma) [21] were identified using a K-means algorithm. Moradi et al. extracted the fractal dimensions of ultrasonic RF signals for the classification of pig livers, chicken breasts, steaks, and bovine livers. The accuracies were in the range of 68%–96%, i.e., higher than the natural split of the data [22]. Six features, including the spectral intercept, slope, and four sums of the amplitude values in four different frequency bands, were used to identify prostate cancer based on a neural network classifier. This method provided a mean accuracy of 91%, sensitivity of 92%, and specificity of 90% [23]. Uniyal et al. extracted ultrasound RF signals, B-mode textures, and attenuation features for support vector machine (SVM) and random forest (RDF) classifiers for the classification of malignant and benign breast lesions. The results based on the area under the curve were 0.86 and 0.81, respectively [24]. Therefore, machine learning algorithms based on multiple features from the analysis of ultrasonic RF signals have shown promising tissue classification performance.

In this study, a machine learning approach is proposed based on an adaptive analysis of RF echo signals from blood, aiming to explore the feasibility of RBC aggregation classification. An adaptive empirical wavelet transform (EWT) [25] is used to decompose ultrasonic RF signals echoed from blood into a series of empirical mode functions (EMFs), from which dominant empirical mode functions (DEMFs) are selected. Six statistical characteristics (mean, variance, median, kurtosis, root mean square (RMS), and skewness) of the locally normalized DEMFs are calculated to form primary feature vectors. The RDF and SVM classifiers are trained with the given feature vectors to obtain prediction models for RBC classification. In the experiments, ultrasonic RF echo signals are acquired in a static condition from five groups of six types of porcine blood samples with a hematocrit of 40% prepared by using plasma dilutions of 75%, 60%, 45%, 30%, 15%, and 0%, respectively (resulting in average numbers of aggregated RBCs of 4.28, 2.72, 2.31, 1.83 1.04, and 1.20, respectively). The optimum parameters and combination of the six statistical characteristics are exhaustively searched for using classifiers with a 10-fold cross-validation strategy. Finally, the RDF and SVM classifiers based on the optimum parameters and feature vectors are tested to evaluate the performance of the RBC aggregation classification.

The remainder of this paper is organized as follows. Section 2 describes the methods including the EWT-based analysis of the ultrasonic RF echo signals, characteristic extraction, and classifiers. The experimental section elaborates on the in vitro experiments performed on the prepared porcine blood samples using the SonixTOUCH RP system. This is followed by the experimental results, discussion, and conclusions.

Section snippets

Methods

Fig. 1 presents an overall classification framework based on an adaptive EWT analysis of ultrasonic RF echo signals. The framework consists of the following steps. In Step 1, RF signals echoed from blood are decomposed into a series of EMFs, from which the DEMFs are selected. In Step 2, six statistical characteristics, including the mean, variance, median, kurtosis, RMS, and skewness of the locally normalized DEMFs, are calculated to form primary feature vectors. The optimum parameters of the

Blood sample preparation

In this study, a total of 4 L of fresh porcine whole blood from 10 adult pigs was obtained from a local slaughterhouse, and was anticoagulated with a 10 g/L heparin injection. Approval was obtained from the Ethics Committee of the Medical School at Yunnan University (approval number: yuncae20180307). The whole blood samples were centrifuged at 2000 g for 15 min at room temperature to separate the plasma and RBCs. To simulate a wide range of aggregation levels, different proportions of the total

Results

Fig. 6 shows B-Mode images for blood samples with different PDs, based on a central frequency of 10 MHz and transmission power level of 0 dB. Fig. 7 shows the microscopic images obtained by the SAGA optical microscope (Shenying Optics Co. Ltd., Suzhou, China). As shown, the RBCs in the microscopic images gradually increase with an increasing PD. In the corresponding B-mode images, the echogenicity values for the aggregated RBCs (PDs of 30%, 45%, 60%, and 75%) are higher than those for the non-

Discussion

As a common pathophysiologic phenomenon, RBC aggregation is an important characteristic of hemorheology, and is closely related to hemodynamics [18]. Under normal physiological conditions, moderate RBC aggregation is necessary to maintain in vivo circulation and/or microcirculation perfusion. In some pathological conditions (e.g., ischemic cardio-cerebrovascular disease and diabetes), excessive RBC aggregation can easily form venous thromboembolisms (deep vein thrombosis and pulmonary

Conclusions

In this study, machine learning based on an adaptive EWT analysis of ultrasonic RF echo signals is proposed for RBC aggregation classification. In this method, the ultrasonic RF signals are decomposed into a series of EMFs with an EWT algorithm, and then the DEMFs containing the most power of the signals are selected. The best combination of variance, kurtosis, and RMS is determined from six statistical characteristics of locally normalized DEMFs to form the optimum feature vectors. Finally,

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This work was supported by Grants (Nos. 81771928 and 61561049) from the National Natural Science Foundation of China, and University Key Lab of Electronic Information Processing of High Altitude Medicine, Yunnan Province.

References (58)

  • B. Neu, H.J. Meiselman, The role of macromolecules in stabilization and de-stabilization of biofluids, in: G.M....
  • M.W. Rampling

    Haemorheology and the inflammatory process

    Clin. Hemorheol. Microcirc.

    (1998)
  • H.J. Meiselman

    Red blood cell aggregation: 45 Years being curious

    Biorheology

    (2009)
  • B. Sigel et al.

    Red cell aggregation as a cause of blood flow echogenicity

    Radiology

    (1983)
  • B. Sigel et al.

    Variable ultrasound echogenic-ity in flowing blood

    Science

    (1982)
  • M. Boynard et al.

    Size determination of red blood cell aggregates induced by dextran using ultrasound backscattering phenomenon

    Biorheology

    (1990)
  • A. Amararene, J. Gennisson, A. Rabhi, G. Cloutier, Quantification of red blood cell aggregation using an ultrasound...
  • F.T.H. Yu et al.

    Experimental ultrasound characterization of red blood cell aggregation using the structure factor size estimator

    J. Acoust. Soc. Am.

    (2007)
  • D. Savéry et al.

    A point process approach to assess the frequency dependence of ultrasound backscattering by aggregating red blood cells

    J. Acoust. Soc. Am.

    (2001)
  • E. Franceschini et al.

    Simultaneous estimation of attenuation and structure parameters of aggregated red blood cells from backscatter measurements

    J. Acoust. Soc. Am.

    (2008)
  • J. Tripette et al.

    In vivo venous assessment of red blood cell aggregate sizes in diabetic patients with a quantitativecellular ultrasound imaging method: Proof of concept

    PLoS ONE

    (2015)
  • R.K. Saha et al.

    Assessment of accuracy of the structure-factor-size-estimator method in determining red blood cell aggregate size from ultrasound spectral backscatter coefficient

    J. Acoust. Soc. Am.

    (2011)
  • E. Franceschini, B. Metzger, G. Cloutier, An effective medium model for ultrasound blood characterization, in: IEEE...
  • C.C. Huang et al.

    Detection of coagulating blood under steady flow by statistical analysis of backscattered signals

    IEEE Trans. Ultrason. Ferroelectr. Freq. Control.

    (2007)
  • C. Yoon

    Spectrum analysis for assessing red blood cell aggregation using high-frequency ultrasound array transducer

    Biochem. Eng. Lett.

    (2017)
  • M. Moradi et al.

    A new approach to analysis of RF ultrasound echo signals for tissue characterization: animal studies

    Proc Spie

    (2007)
  • M. Moradi et al.

    Discrete Fourier analysis of ultrasound RF time series for detection of prostate cancer

    IEEE EMBC

    (2007)
  • N. Uniyal et al.

    Ultrasound RF time series for classification of breast lesions

    IEEE Trans. Med. Imaging

    (2014)
  • J. Gilles

    Empirical wavelet transform

    IEEE Trans. Signal Process.

    (2013)
  • Cited by (9)

    • Osteoporosis Diagnostic Model Using a Multichannel Convolutional Neural Network Based on Quantitative Ultrasound Radiofrequency Signal

      2022, Ultrasound in Medicine and Biology
      Citation Excerpt :

      The RF signal is a native unfiltered ultrasound signal that contains a wealth of information on tissues through which ultrasonic waves pass (Shan et al. 2015; Uniyal et al. 2015; Granchi et al. 2016; Liu et al. 2019a). The most common analysis strategy for RF signal in other studies was to extract the quantitative parameters from the waveform, such as the effective acoustic concentration, the average of the frequency spectrum and the kurtosis of the amplitude (Uniyal et al. 2015; Shen et al. 2019; Liao et al. 2021). These quantitative parameters are handcrafted features and may not reflect different categories of waveforms with subtle differences (Liu et al. 2019b).

    • Red Blood Cell Aggregation Classification Based on Ultrasonic Radiofrequency Echo Signals by An Improved Convolutional Neural Network

      2022, 2022 3rd International Conference on Computer Vision, Image and Deep Learning and International Conference on Computer Engineering and Applications, CVIDL and ICCEA 2022
    View all citing articles on Scopus
    View full text