Multi-Parameter Sub-Hertz Analysis of Viscoelasticity with a Quality Metric for Differentiation of Breast Masses

Abstract

We applied sub-Hertz analysis of viscoelasticity (SAVE) to differentiate breast masses in pre-biopsy patients. Tissue response during external ramp-and-hold stress was ultrasonically detected. Displacements were used to acquire tissue viscoelastic parameters. The fast instantaneous response and slow creep-like deformations were modeled as the response of a linear standard solid from which viscoelastic parameters were estimated. These parameters were used in a multivariable classification framework to differentiate malignant from benign masses identified by pathology. When employing all viscoelasticity parameters, SAVE resulted in 71.43% accuracy in differentiating lesions. When combined with ultrasound features and lesion size, accuracy was 82.24%. Adding a quality metric based on uniaxial motion increased the accuracy to 81.25%. When all three were combined with SAVE, accuracy was 91.3%. These results confirm the utility of SAVE as a robust ultrasound-based diagnostic tool for noninvasive differentiation of breast masses when used as stand-alone biomarkers or in conjunction with ultrasonic features.

INTRODUCTION

Normal human breast tissue consists of cellular and non-cellular components that regulate the physical structure and functionality of this organ. The growth of abnormalities impacts the mechanobiology of the microenvironment at the lesion site, leading to macroscopic changes in the tissue mechanical properties (Dvorak, et al. 2011). Conventionally, malignant tumors have been associated with high stiffness, and if superficial enough, can often be physically sensed with palpation (Klein 2005). However, deep-seated masses can be hard to identify using palpation. Hence, imaging of the tissue elastic properties, elastography, has emerged as a new imaging modality to overcome this problem by providing information about the tissue internal stiffness distribution noninvasively (Garra, et al. 1997, Hiltawsky, et al. 2001, Sinkus, et al. 2000).

To differentiate between malignant and benign tumors, a large body of research has been focused on the measurement of tumor elasticity. When using ultrasound as the imaging modality, these methods monitor tissue deformations in response to external or internal forces and provide maps that represent stiffness differences either qualitatively (e.g. strain elastography) (Barr, et al. 2012, Garra, et al. 1997, Hiltawsky, et al. 2001) or quantitatively (e.g., shear wave elastography) (Barr, et al. 2012, Bayat, et al. 2017, Denis, et al. 2015, Evans, et al. 2010, Ghosh, et al. 2017).

An important aspect of the tissue is its biphasic nature that stems from its large water content. The interaction of the interstitial fluid with the porous-like environment of the solid matrix mimics the behavior of a viscoelastic solid material at macroscopic scales. Instead of instantaneous response to a sudden stress, such materials present retardations as a portion of the applied stress is dissipated in the form of frictional forces. A comprehensive model that can relate the microscopic alterations to macroscopically observable variations, in terms of rheological responses, remains the subject of many ongoing studies (Balleyguier, et al. 2018, Shi, et al. 2018, Zhang, et al. 2018). Nevertheless, the additional information provided by a simplified elastic solid model (i.e., the underlying basis of the elastography methods) has shown to greatly enhance the diagnostic accuracy of the ultrasound breast examination (Berg, et al. 2012). Hence, as a natural extension, viscoelastic models can be considered that include the elastic model as a special case and may provide additional useful information regarding the tissue state.

Among other factors, the choice of an appropriate viscoelastic model for analyzing tissue behavior depends on both size and time scales. The former is mainly constrained by the imaging system resolution beyond which only a macroscopic view of the tissue response can be obtained, and the latter is determined by the feasibility of a test for in vivo evaluation. While classical rheological models allow responses that may be represented as the superposition of an infinite number of modes (e.g., Prony series), the choice of a reasonable time scale not only depends on the presumed complexity of the model but also on the measurement constraints imposed by the experimental setup. For example, respiration-induced motions can be a significant source of error when considering long-term responses that last many seconds, minutes, or hours. Hence, the useful duration for which reliable parameters can be obtained can be limited by the maximum time that patients can hold their breath (e.g., less than 10 seconds) (Bayat, et al. 2018) . Another crucial component of viscoelasticity testing is the type of excitation. For in vivo studies, Food and Drug Administration (FDA) requirements limit the maximum thermal or mechanical stress that can be imposed upon the tissue. These requirements, in turn, limit the utility of methods that depend on the creation of internal forces, for example, via acoustic radiation (Amador, et al. 2012). Mechanical excitation via external forces provides a reliable approach for completing a creep-like test that can be performed in vivo (Qiu, et al. 2008, Sridhar, et al. 2007). The preliminary study in (Qiu, et al. 2008) showed that, among other parameters, the introduced viscoelastic time constant, T₁, was able to effectively differentiate between benign and malignant lesions, thereby providing a new potential biomarker for differentiation of breast masses. With the aid of an automated external compression device integrated with ultrasound imaging, the study in (Bayat, et al. 2018) provided a robust and automated mechanism to extract viscoelastic parameters at sub-Hertz frequencies in a tune-free and noninvasive fashion. In the current study, sub-Hertz analysis of viscoelasticity (SAVE) was performed on a large group of patients undergoing biopsy. The main aim of this study was to evaluate the diagnostic performance of SAVE parameters when used as individual biomarkers, as well as their classification power when used in a multivariable fashion and in conjunction with other ultrasound imaging metrics.

MATERIAL AND METHODS

Automated ramp-and-hold creep-like test

A custom-made compression device integrated with a reconfigurable ultrasound machine (Verasonics, Kirkland, WA) was used to simultaneously induce a uniaxial compression and image the internal deformations via high frame-rate imaging using a linear array (L11–4v, Verasonics, Kirkland, WA), as explained in (Nabavizadeh, et al. 2017). The compression device was set to apply a ramp-and-hold force on the breast with the ramp speed of 8 N/sec and the constant force of 2 N. The ramp time was, therefore, 250 msec, and the hold time was approximately 10 sec. The compression plate had a surface area of 40 × 60 mm. Minimal initial compression was applied to ensure proper contact of the compression plate with breast skin while ensuring a linear tissue response. To decrease inter-frame decorrelation, the ultrasound system acquired high frame plane-wave data in in-phase and quadrature (IQ) format at 200 frames per sec. A two-dimensional autocorrelation method (Loupas, et al. 1995) was used to calculate the displacement data. The frame-to-frame displacement data were properly accumulated to result in the last-to-first frame (Lagrangian) displacement. Only axial displacement was considered. A band-limited staggered strain filter (similar to (Srinivasan, et al. 2002)) was then applied on the accumulated displacement to obtain accumulated strain temporal curves for all grid points in the imaging field of view.

Motion-compensated cross-correlation

Depending on the level of bonding to the surrounding tissue, a lesion under compression might present some level of translational motion in the lateral and elevational directions. These unwanted motions may adversely impact the quality of the estimated viscoelastic parameters when a uniaxial creep-like test is assumed. Hence, to assess the quality of uniaxial testing, using the estimated displacement field, each frame was stretched back to its original location at the pre-compression state. These motion-compensated frames were then correlated with the pre-compression frame to obtain a quality metric (QM) as a function of the frame index. For the i^th frame, the QM(i) was calculated as

$$QM\left(i\right)=\frac{{\sum }_{mϵ\mathit{\text{axial}}\mathit{\text{grid}}}{\sum }_{nϵ\mathit{\text{lateral}}\mathit{\text{grid}}}I{Q}_0\left(m,n\right)\overline{\widehat{I{Q}_{\iota }}\left(m,n\right)}}{\sqrt{{\sum }_{mϵ\mathit{\text{axial}}\mathit{\text{grid}}}{\sum }_{nϵ\mathit{\text{lateral}}\mathit{\text{grid}}}{|I{Q}_i\left(m,n\right)|}^2}\sqrt{{\sum }_{mϵ\mathit{\text{axial}}\mathit{\text{grid}}}{\sum }_{nϵ\mathit{\text{lateral}}\mathit{\text{grid}}}{|\widehat{I{Q}_i}\left(m,n\right)|}^2}}$$ Eq 1

where IQ₀(m, n) is the IQ data obtained from inside the lesion area at the pre-compressed state, and $\widehat{I{Q}_i}\left(m,n\right)$ is the motion-compensated IQ data obtained from the i^th frame during the compression. The horizontal bar represents complex conjugation. Since a unidirectional speckle tracking was utilized, only uniaxial lesion motion should result in high QM values. Hence, for each test on the lesion, QM was obtained and used as a measure of the test uniaxial reliability.

Viscoelasticity model

A standard linear solid model was used to acquire viscoelasticity parameters using a two-step fitting approach explained in (Bayat, et al. 2018). Briefly, the initial instantaneous response was approximated as fully elastic and was modelled using a purely elastic component with stiffness E₀. This parameter was estimated as $E_0=\frac{{\sigma }_0}{{ϵ}_0}$, where σ₀ represents the total stress obtained from the load cell reading and knowledge of the compression plate surface area, and ϵ₀ represents the initial strain jump as estimated ultrasonically from the displacement data. With appropriate assumptions regarding substantial difference in rate of changes of the applied stress and measured strain (Bayat, et al. 2018), the remaining part of the strain, ϵ(t) , was considered as the response of a first-order Kelvin-Voigt solid and it was related to the input stress σ₀ as

$$ϵ\left(t\right)=\frac{{\sigma }_0}{E_1}\left(1-e^{-t/T_1}\right)$$ Eq 2

where E₁ and T₁ represent the creep elasticity and viscoelastic retardation time, respectively. Levenberg-Marquardt nonlinear least-square fitting (Moré 1978) was used in MATLAB (Mathworks, Natick, MA) to perform the fitting and acquire the viscoelasticity parameters E₁ and T₁. Normalized fitting error was calculated as

$$\overline{\chi }=\sqrt{\int {\left({ϵ}_m\left(t\right)-ϵ\left(t\right)\right)}^{2}dt/\sqrt{\int {ϵ}_m^2\left(t\right)dt}}\times 100$$ Eq 3

where ϵ_m(t), ϵ(t) represent the measured and model-predicted strain profiles respectively after bandwidth reduction to exclude zero mean cyclic model deviations induced by cardiac activity as explained before (Bayat, et al. 2018).

Patient population and scanning

Pre-biopsy patients with at least one suspicious breast mass were recruited according to an approved Mayo Clinic Institutional Review Board (IRB) protocol. Written consent was obtained from each patient prior to the study. Patients were scanned in supine. The lesion was first identified using B-mode ultrasound. Probe orientation was adjusted such that the compression plate could compress the breast against the chest wall primarily in the axial direction with minimal lateral motion. Multiple acquisitions were obtained from each patient to ensure consistency. Patients were asked to hold their breath during each acquisition, lasting around 10 seconds. All acquisitions were used for subsequent analyses.

Region of interest (ROI) selection

The automated ROI selection method in (Bayat, et al. 2018) was used to obtain different viscoelasticity parameters of the lesion site and the surrounding non-lesion tissue. A lesion quantification domain was obtained from the pre-compressed B-mode images. To obtain viscoelasticity parameters of the surrounding non-lesion tissue, a dilation factor of 100% was used on the lesion boundary. To overcome the uncertainty associated with determination of the true lesion margin from the B-mode images (a known characteristic of mostly malignant lesions), a leave-out area was created using a 30% dilation of the lesion boundary. The lesion and background values for a viscoelastic parameter X were obtained as X_lesion and X_nonlesion for lesion and surrounding non-lesion tissue, respectively. For each parameter, a contrast value was also defined as

$$\mathit{\text{contrast}}\left(X\right)=2\frac{X_{\mathit{\text{lesion}}}-X_{\mathit{\text{nonlesion}}}}{X_{\mathit{\text{lesion}}}+X_{\mathit{\text{nonlesion}}}}\text{'}$$ Eq 4

Statistical analysis

All statistical analyses were performed in MedCalc (MedCalc Software bvba, Ostend, Belgium). For each patient, statistical measures of viscoelastic parameters were obtained in terms of mean and standard deviation. Statistical averaging was performed to obtain one measure from multiple acquisitions from each patient. The Student t-test was used to test for significant differences. To study the advantage of using QM as a test reliability factor, all analyzes were repeated by including only those cases with sustained QM > 0.1 over the entire deformation time. This threshold was selected empirically based on visual inspection of the lesion deformations in cases with and without apparent nonaxial motions and deformations. Error bar plot of mean value distribution across the relevant cohorts was created for each quantitative viscoelastic metric with 95% confident interval vertical bars representing the accuracy of estimation. Paired t-test with p-value was used to analyze significance of differences; p<0.05 considered statistically significant. Quantitative values reported in text were reported mean±standard deviation.

RESULTS

The SAVE method was performed on 161 pre-biopsy patients in vivo. Only one lesion was studied in patients with multiple lesions. Based on breast imaging-reporting and data system (BIRADS), lesions consisted of 40 BIRADS 5, 108 BIRADS 4, and 13 BIRADS 3. From this group, pathology revealed 77 cases as malignant and 84 cases as benign. The BIRADS 3 group were considered benign based on either follow up visits or clinical examinations without a biopsy.

The median patient age was 56 years, and the mean lesion size ± standard deviation was 16.63 ± 9.38 mm.

Instantaneous elasticity E₀

Representative maps of the instantaneous elasticity, E₀, acquired from the fast portion of the tissue response to the external ramp-and hold stress, are shown for three benign and three malignant cases in Figure 2 and Figure 3, respectively. The E₀ reconstruction maps are noticeably different between the benign and malignant cases. While benign cases appear to have instantaneous elasticity similar to the surrounding tissue, malignant cases present a conspicuous contrast. Additionally, the boundary of the lesion determined from the B-mode images favorably matches the elevated E₀ in malignant cases. Furthermore, malignant cases presented higher E₀ reconstruction heterogeneity on the lesion site compared to the benign lesions, as seen in Figure 3. The mean ± standard deviation of E₀ on the lesion site for the benign cases was 15.84 ± 15.64 kPa, while for the malignant cases it was 21.91 ± 13.77 kPa, which was significantly higher than that of the benign (p=0.01, Figure 4a). Malignant cases presented higher E₀ contrast (0.42 ± 0.37) compared to the benign lesions (0.17 ± 0.33), and these values were significantly different (p<0.0001) (Figure 4b). In all images, areas with large fitting errors were excluded from the quantification, as explained in (Bayat, et al. 2018). Fitting erorr maps indicated good agreement between oberserved strain data and fitted model with vlaues mostly less than 10% for small and shallow lesions and less than 20% for large and deep seated lesions (Figure 2 and Figure 3). Large lesions presented higher fitting errors on the lesion site compared to small lesions, which may be suggestive of underlying heterogeneity.

Figure 2:

B-mode image and viscoelasticity parameters obtained by SAVE from three representative cases that were revealed as benign by pathology. Each column represents a separate case, and each row represents a specific viscoelastic parameter or B-mode image. The specific subtypes of these cases were: Case 1: benign sclerosing adenosis with the apocrine change, Case 2: fibroadenoma and Case 3: benign clustered apocrine cysts. The minimum QM values for case 1,2 and 3 were 0.05, 0.15 and 0.12, respectively.

Figure 3:

B-mode image and viscoelasticity parameters obtained by SAVE from three representative cases that were revealed as malignant by pathology. Each column represents a separate case, and each row represents a specific viscoelastic parameter or B-mode image. The specific subtypes of these cases were: invasive mammary carcinoma with mixed ductal and lobular features, grade II, calcifications present in invasive carcinoma, Case 2: invasive mammary carcinoma with abundant tumor-infiltrating lymphocytes, grade III and Case 3: Invasive Lobular Carcinoma (ILC) Grade II. The minimum QM values for case 1,2 and 3 were 0.48, 0.05 and 0.07, respectively.

Figure 4:

(a,c,e) Error-bar plot of different viscoelasticity parameters obtained by SAVE for lesion (red) and non-lesion (blue) tissue in benign and malignant lesions; (b,d,f) error-bar plot of corresponding contrast values in benign and malignant lesions. In each plot, middle horizontal bar represents mean and extended error bars represent 95% confidence interval for the corresponding parameter. p-values for paired parameters are shown with asterisk indicating significant, i.e. p<0.05.

Creep elasticity E₁

As seen in Figure 2, two of the benign cases did not present noticeable differences in E₁ between lesion and non-lesion areas. In contrast, the representative malignant cases presented high E₁ values on the lesion with considerable reconstruction heterogeneity (Figure 3). The mean E₁ in the benign group was 55.60 ± 32.73 kPa. This value was 86.68 ± 61.05 kPa in the malignant group and was significantly higher than in benign cases (p=0.0001, Figure 4c). When considering E₁ contrast values, malignant cases showed a significantly higher E₁ contrast, 0.37 ± 0.32, compared to the benign cases which was 0.12 ± 0.25 (p<0.0001, Figure 4d).

Viscoelastic retardation time T₁

Viscoelastic retardation time for the representative benign and malignant cases can be seen in Figure 2 and Figure 3, respectively. All three benign lesions showed elevated T₁ values on the lesion area, though with different levels of heterogeneity. Two of the malignant cases showed noticeably lower T₁ values on the lesion area compared to the surrounding non-lesion tissue. When testing for statistical differences, the mean T₁ value in benign lesions was 1.66±0.7s and it was 1.60±0.65s in the malignant group (p=0.57). Benign lesions had a slightly higher T₁ value (1.66±0.7s) compared to the surrounding non-lesion area (mean 1.63±0.66s) (p=0.74, Figure 4e). The T₁ value of the malignant lesions (1.60±0.65s) was slightly lower than the surrounding non-lesion area (1.64±0.58s) (p=0.7, Figure 4e). The T₁ contrast in the benign lesions was mostly positive with an average of 0.0048±0.15 (Figure 4f). However, the T₁ contrast in the malignant lesions was mostly negative with an average of −0.047±0.19 (Figure 4f). The difference in the two groups of malignant and benign lesions was not found to be statistically significant (p=0.056).

QM Assessment

Figure 5 shows the 2-D maps of normalized fitting error, $\overline{\chi }$, in three repeated acquisitions from a single representative malignant lesion. Strong similarities can be observed in error maps, which is indicative of the repeatability of our tests. However, acquisition 3 represents a noticeably lower fitting error with fewer fluctuations throughout the entire reconstruction map compared to the first two acquisitions. These findings coincide with the observation that the QM value for acquisition 3 consistently remained above the predefined threshold of 0.1, while the opposite was true for acquisitions 1 and 2.

Figure 5:

2-D normalized fitting error maps in three repeated acquisitions from a single lesion and corresponding QM curves as a function of frame number. Only the third acquisition exceeded the predefined threshold (Thr = 0.1) for QM with visible reduced fitting errors compared to the first two acquisitions. The mean $\overline{\text{χ}}$ for Acquisition 1, 2 and 3 were 4.12%, 4.16% and 2.82% respectively.

Receiver Operator Curve (ROC) analysis

The ROC curves obtained based on the data from 161 patients when using viscoelastic parameters E₀, E₁, and T₁ to classify breast lesions in terms of mean value on the lesion, mean value of the contrast, and standard deviation on the lesion are shown in Figure 6a, b, and c, respectively. When considering mean values in the lesion area, elasticity values E₀ and E₁ provided better classification performance in terms of area under the ROC curve (AUC) compared to the retardation time, T₁. When considering contrast values instead of mean values on the lesion area, the classification performance based on T₁ showed improvement, though still inferior to those of E₀ and E₁. The standard deviation of the parameters on the lesion area showed similar classification performance for all three viscoelastic parameters, with elasticity parameters, E₀ and E₁, showing slightly better classification compared to T₁ (Figure 6c).

Figure 6:

Receiver operator curves when using the (a) mean value of different viscoelastic parameters on the lesion area for classification, (b) when considering contrast values based on different viscoelastic parameters for classification, and (c) when using the standard deviation of different viscoelastic parameters on the lesion area for classification.

ROC analysis with QM

The results of ROC analysis concerning the 48 cases (N=20 malignant, N=26 benign) that exceeded the required QM are presented in Figure 7. For each case, at least one acquisition provided a QM above the threshold. When considering the mean value of the parameters on the lesion area, elastic parameters, E₀ and E₁, had considerably better classification performances than T₁. The contrast and standard deviation values on the lesion area, though, performed similarly for different viscoelastic parameters, as seen in Figure 7b and c, respectively.

Figure 7:

Receiver operator curves when using (a) the mean value of different viscoelastic parameters on the lesion area for classification of cases that exceeded the quality metric (QM) requirement, (b) when considering contrast values based on different viscoelastic parameters for classification of cases that exceeded the QM requirement, (c) when using the standard deviation of different viscoelastic parameters on the lesion area for classification of cases that exceeded the QM requirement.

Multivariable Lesion Classification Using SAVE Parameters

The outcomes of using a logistic regression model for the classification of the breast masses when using all SAVE parameters simultaneously are presented in Table 1. The regression model was able to identify 63 of 84 benign lesions correctly, resulting in a specificity value of 75%. The model also identified 52 of 77 malignant lesions correctly, resulting in a sensitivity value of 67.53%. The overall classification accuracy of this model was 71.43%, with an AUC of 0.78.

Table 1:

Logistic regression analysis using all SAVE parameters in conjunction with ultrasound imaging information and quality metric

	Logistic regression variables
	Average values of E0, E1, T1 on lesion Standard deviation of E0, E1, T1 on lesion Contrast value of E0, E1, T1		Average values of E0, E1, T1 on lesion Standard deviation E0, E1, T1 of on lesion Contrast value of E0, E1, T1 BIRADS Lesion size
	No QM	QM	No QM	QM
Malignant	N= 77	N= 21	N= 72	N= 20
Benign	N= 84	N= 27	N= 80	N= 26
Sensitivity	67.53%	66.67%	77.78%	80.00%
Specificity	75.00%	92.59%	86.25%	100.00%
Specificity	75.00%	92.59%	86.25%	100.00%
Accuracy	71.43%	81.25%	82.24%	91.30%
AUC	0.78 (CI: 0.71–0.84)	0.85 (CI: 0.71–0.93)	0.89 (CI: 0.83–0.94)	0.93 (CI: 0.81–0.98)

AUC: area under the curve, CI: confidence interval, E₀ instantaneous elasticity, E₁ creep elasticity, T₁ retardation time, QM quality metric

A logistic regression classification model was utilized to classify lesions using only SAVE acquisitions that exceeded the QM requirement. The results are also summarized in Table 1. The addition of QM resulted in similar sensitivity (66.67%) but significantly increased the specificity by 17.59% and overall accuracy by 9.82%. The AUC was also increased to 0.85.

Multivariable Lesion Classification Using SAVE Parameters and Ultrasound Imaging Assessment

Sub-Hertz analysis of viscoelasticity parameters in combination with ultrasound imaging data, including the BIRADS and lesion size, were used in a multivariable logistic regression for classification of lesions. The addition of BIRADS and lesion size to viscoelasticity data increased sensitivity, specificity, and overall accuracy by 10.25%, 11.25%, and 10.81%, respectively. The AUC was also improved by 0.11. This analysis was repeated for cases that exceeded the QM requirement. The addition of the QM increased the sensitivity to 80%, specificity to 100%, and overall accuracy to 91.3%. The AUC was also increased to 0.93. These results are summarized in Table 1.

DISCUSSION

Viscoelastic properties of tissue may be indicative of substantial physiological or pathological changes. In this study, we presented the utility of a low-frequency viscoelasticity test for the classification of breast masses. The method mimicked a ramp-and-hold viscoelastic creep test from which different parameters, such as instantaneous and creep elasticity and retardation time, were obtained. Using an automated compression device integrated with an ultrasound machine, SAVE was tested in a large group of patients, and the results were analyzed in comparison to the pathology results as the gold standard. The representative imaging results highlighted a distinctly elevated lesion elasticity compared to the surrounding tissue in the malignant lesions compared to the benign. The T₁ values were mostly higher in the benign lesions compared to surrounding tissue, while the opposite trend was observed in the malignant lesions. This is in agreement with the results in (Qiu, et al. 2008), in which only non-palpable lesions were studied. The current study included a more diverse group of breast lesions, resulting in a more realistic sampling of T₁ behavior in a large group of prebiopsy patients. The viscoelasticity parameters E₁ and T₁ showed considerably higher reconstruction heterogeneity compared to the instantaneous elasticity, E₀. Hence, the standard deviation of the parameters obtained from the lesion area, considered as a surrogate for heterogeneity, resulted in classification results similar to those of average values on the lesion. This observation is in agreement with other investigators who have found heterogeneity as a hallmark of lesion malignancy (Liu, et al. 2015). When considering viscoelasticity parameters individually, elasticity parameters obtained from the lesion area provided better classification compared to the time constant. This is in concordance with the results of a previous study (Bayat, et al. 2018) which reported that intrinsic time constant values may be adversely affected by the geometry and boundary conditions. However, when considering imaging contrast values, all viscoelasticity parameters performed similarly, though each parameter presented moderate differentiation capability when used for the classification of malignant and benign lesions.

The results of the multivariable analysis showed that lesion classification was significantly improved when considering all parameters simultaneously. While non-palpable malignant lesions (with low elasticity contrast) have shown to present mostly negative T₁ contrast (Qiu, et al. 2008), highly palpable malignant lesions (with high elasticity contrast) may present different T₁ features. This further justifies the improved performance of the multivariable classifier compared to a single parameter (see supplementary material). Ultrasound information alone, which is summarized in BIRADS lexicon can improve breast lesion classifiers based on elasticity (Bayat, et al. 2017). However, it carries low specificity in determination of lesion malignancy. The multicenter study by (Berg, et al. 2012) reported a specificity of 61.1% among a diverse cohort of breast lesions that included 102 BIRADS 2, 303 BIRADS 3, 347 BIRADS 4 and 187 BIRADS 5. In our study, the viscoelastic parameters obtained by SAVE provided a specificity of 75% and 92.59% without and with QM, respectively. The combination of ultrasound BIRADS data and SAVE viscoelasticity parameters significantly improved specificity to 86.25% and 100% without and with QM, respectively, indicating the synergistic value of combined modalities in the enhanced characterization of the benign masses that can assist in reducing unnecessary biopsies. Our analysis comprised only 13 BIRADS 3 and no BIRADS 2; SAVE is expected to be even more specific in a balanced cohort of patients consisting of more benign cases.

In malignant lesions, lesion size may be considered as a potential biomarker of tumor aggressiveness and stage (Denis, et al. 2016). Hence, addition of these parameters considerably improved lesion classification when combined with the SAVE parameters.

It is known that in vivo application of compression elastography, especially if done manually, is prone to errors and artifacts due to heterogeneity in composition, location, and geometry of breast lesions. These features are hard to replicate in experimental models, so to evaluate QM, in this study we resorted to an empirical approach where we could quantitatively assess the quality of elastograms over time in real patients. The addition of a data acquisition QM improved the performance of lesion classification. Lesion mobility may be a challenging factor in the successful utilization of SAVE. The motion-compensated cross-correlation metric provided a versatile mechanism to evaluate the goodness of uniaxial lesion compression, which is essential in the successful implementation of SAVE. In this study, QM was obtained and applied retrospectively and a threshold of 0.1 was chosen empirically, the highest value without underpowering statistical analysis due to cohort size reduction. However, the true value of QM will be while applying the SAVE tests to ensure more reliable data collection via repeated tests. Such approaches have also been utilized in shear wave elastography to provide a means of feedback to improve clinical utility (Barr and Zhang 2014). Future studies will focus on optimizing QM and studying its utility in obtaining high quality data, ideally in a real-time manner, to repeat an acquisition if needed.

In addition to clinical value, the methods presented in this study provided a means to cover the unexplored dynamic range between quasi-static elastography and dynamic methods using 10–1000 Hz excitation (e.g., magnetic resonance and acoustic radiation force elastography) and may open a new way for understanding the mechanobiological manifestation of breast tissue alterations due to different pathologies. Given the versatility and noninvasive nature of SAVE, this, in turn, can assist in improving the evaluation of prognosis by timely monitoring of tissue alterations in response to chemotherapeutic medications and radiotherapy.

CONCLUSION

Viscoelasticity parameters obtained by the SAVE method, when combined with other ultrasound data, can provide a robust noninvasive tool for differentiation of breast lesions.

Figure 1:

Schematic diagram showing steps for calculating the quality metric (QM). The red closed trajectory represents the hypothetic lesion margin in the pre-compressed state. Blue closed trajectory shows the lesion margin during compression. Slight boundary mismatch represents displacement tracking and interpolation errors associated with motion compensation. The dashed line represents a threshold to determine reliable tracking based on the QM.