Low-cost quasi-real-time elastography using B-mode ultrasound...

Bio-Medical Materials and Engineering 24 (2014) 1673–1692 1673DOI 10.3233/BME-140980IOS Press

Low-cost quasi-real-time elastography usingB-mode ultrasound images

Hyock-Ju Kwon ∗ and Jiwon Lee

Department of Mechanical and Mechatronics Engineering, University of Waterloo, Waterloo, ON,Canada

Received 18 June 2013Accepted 13 December 2013

Abstract. A low cost, quasi real-time elastography system, displacement-gradient elastography (DGE), was developed byapplying digital image correlation (DIC) method and smoothing algorithm to B-mode ultrasound images. In order to achievequasi real-time elastogram display, a new fast pattern matching algorithm, decoupled cross-correlation (DCC), was proposedand validated. By applying the DGE to various phantoms, elastograms were generated to identify the lesion with wide variationsof stiffness ratio and applied strain. The performance of DGE was qualitatively compared with those from a high-end ultrasoundscanner using the elastograms of a commercial elastography breast phantom. DGE was also applied to the ultrasound imagesof human breast lesions in various BI-RADS categories. This study suggests that DGE may have comparable performance toconventional elastography in detecting breast cancer, while it can be easily implemented onto conventional ultrasound scanners.

Keywords: Displacement-gradient elastography, elastogram, digital image correlation, breast cancer

1. Introduction

Breast cancer in women is a major health problem throughout the world. It is the most common typeof cancer among women in both developed and developing countries [1,2]. More than 1.1 million casesare diagnosed and more than 600,000 patients die of it worldwide each year [2]. Over the last decade,researchers have strived to develop and improve early detection tools such as mammography, ultrasound(US) and MRI [3–7]. However, none of these methods, whether applied individually or in combination,has proved universally effective [5–7]. In particular, BI-RADS (Breast Imaging-Reporting and DataSystem) category 3 and 4 lesions are notoriously difficult to characterize with these methods, thus 70%to 90% of biopsies are performed for benign lesions in these categories [8].

Recently, elastography has emerged as a new modality to detect and classify pathological lesions us-ing mechanical strains [9–12]. Since pathological lesions are normally stiffer than the normal tissue,the strain in the lesion is less than the surrounding tissue under compression [13,14]. Strains are cal-

*Address for correspondence: Hyock-Ju Kwon, Department of Mechanical and Mechatronics Engineering, University ofWaterloo, 200 University Ave. W., Waterloo, ON, N2L 3G1, Canada. Tel.: +1 519 888 4567, ext: 33427, Fax: +1 519 885 5862;E-mail: [email protected].

0959-2989/14/$27.50 © 2014 – IOS Press and the authors. All rights reserved

1674 H.-J. Kwon and J. Lee / Low-cost quasi-real-time elastography using B-mode ultrasound images

Fig. 1. Schematics of the principles of elastography: (a) time-gradient elastography and (b) displacement-gradient elastography.(Colors are visible in the online version of the article; http://dx.doi.org/10.3233/BME-140980.)

culated using the time-gradient of radiofrequency (RF) echo signals, obtained before and after a slightcompression of the tissue (Fig. 1(a)), i.e.

ε1 =(t1b − t1a) − (t2b − t2a)

t1b − t1a, (1)

where t1a and t1b are the arrival times of the pre-compression echoes from the two reference windows(proximal and distal), respectively, and t2a and t2b are the arrival times of the post-compression echoesfrom the same windows, respectively [12]. Resulting strains are displayed as an image, called elastogram[15–17]. Modern high-end US scanners are implemented with elastography function (e.g. Hitachi HV900; Philips xMATRIX) which can display the color-coded strain information superimposed on B-modeUS images.

However, not many hospitals are equipped with such high-end elastography scanners, particularlyin developing countries. Implementing elastography function onto conventional US scanner is almostimpossible, because RF data cannot be acquired from the US scanner without significant hardware mod-ification [18]. Although some research groups have developed their own dedicated scanners that deliverthe RF data, they cannot be used for clinical applications, without the stringent approval from healthadministration (e.g, Humanitarian Use Device (HUD) approval from FDA in USA).

In this study, we implemented a low-cost elastography system, displacement-gradient elastography(DGE), onto conventional affordable US scanner by acquiring US B-mode images from the US scannerand applying the digital image correlation (DIC) method to the B-mode images. Strains are estimatedfrom the displacements (Fig. 1(b)) by taking the gradients as:

ε1 =(x1b − x1a) − (x2b − x2a)

x1b − x1a, ε2 =

(y1b − y1a) − (y2b − y2a)y1b − y1a

, (2)

H.-J. Kwon and J. Lee / Low-cost quasi-real-time elastography using B-mode ultrasound images 1675

where (x1a, y1a) and (x1b, y1b) are the coordinates of the proximal and distal windows in the pre-compression image, respectively, and (x2a, y2a) and (x2b, y2b) are the coordinates of the same windows inthe post-compression image. Although B-mode images contain much less information than the RF data,they are readily available from most of the US scanners, thus the implementation can be easily donewithout hardware modification. The major drawbacks of DGE might be the heavy computational load inestimating displacement field using the conventional pattern matching algorithm, thus real-time displayof elastograms was hard to be achieved. In this study, the computational load could be greatly reducedby employing a new pattern matching algorithm, decoupled cross-correlation (DCC) algorithm, to real-ize quasi real-time elastogram display. The performance and the abilities of DGE employing DCC wereevaluated using the custom-built phantoms. DGE was also applied to commercial US breast phantom toqualitatively compare the DGE elastogram with the time-gradient elastogram generated by high-end USscanner. Finally DGE was applied to the human breast US images in different BI-RADS categories toassess the feasibility of DGE in clinical practice.

2. Materials and methods

2.1. Displacement-gradient elastography (DGE) system

2.1.1. System configurationThe schematic of DGE system is illustrated in Fig. 2. US images were taken by the conventional

medical US scanner (Accuvix XQ, Medison) while breast phantom was compressed by the probe orexternal loading device. Video signals from the US scanner were captured by video capture device(DVD EZMaker USB Gold, AverMedia) and recorded on the computer. Most of the US scanners haveexternal video signal output; otherwise, VGA signal to the monitor can be split (e.g., VGA splitter,Belkin) and captured (e.g., VGA2USB external frame grabber, Epiphan). Note that even the low-endvideo capture devices (<$300) provide the resolution (640×480 pixels) and the frame rate (30 fps) highenough to implement the elastography function. In the current system, US images of 560 × 440 pixels(spatial resolution of 81 µm/pixel) were recorded at the frame rate of 5 fps. Synchronization betweenthe captured images and the data from external loading device was conducted by a custom-developedLabview program (National Instruments).

Fig. 2. Schematic of displacement-gradient elastography (DGE) system. (Colors are visible in the online version of the article;http://dx.doi.org/10.3233/BME-140980.)


2.1.2. Decoupled cross-correlation (DCC) algorithmDigital image correlation (DIC) is an image processing method which quantifies the similarity of two

images, utilizing pattern matching algorithm such as fast normalized cross-correlation (FNCC) algorithm[19–21]. Pattern matching is performed by moving the search template (f (x, y) in Fig. 3(a)) in thereference image (Imt0) across the search window (g(x, y)) in the object image (Imt0+Δt) to find thenew position of the template in g(x, y). Mathematically, this involves calculating the normalized cross-correlation (NCC) coefficient γ at each pixel position (u, v) within the search window:

γ(u, v) =

∑x,y(f (x, y) − f̄ )(t(x− u, y − v) − t̄)

{∑

x,y(f (x, y) − f̄ )2∑

x,y(t(x− u, y − v) − t̄)2}1/2, (3)

which yields a value of 1 when two images are exactly matched and a value of 0 when no match is made.For example, consider the process of finding the position of a 2 × 2 template f (x, y) in a 4 × 4 searchwindow g(x, y) in Fig. 3(a). The template is shifted into nine different positions, where at each position,

Fig. 3. Schematics of (a) digital image cross-correlation (DIC) algorithm, and (b) decoupled cross-correlation (DCC) algorithm.(Colors are visible in the online version of the article; http://dx.doi.org/10.3233/BME-140980.)


γ(u, v) is calculated, and the pixel location (u, v) yielding the maximum γmax is regarded as the positionof the template in Imt0+Δt. Note that computational load increases very rapidly as the sizes of templateand search window increase, which is a major obstacle for real-time processing.

In order to reduce the computational load, decoupled cross-correlation (DCC) algorithm was custom-developed for this study, based on the clinical practice of elastography, i.e. breast is compressed by theUS probe to induce the major deformation in the axial (loading) direction, while lateral deformation alsohappens due to Poisson expansion. Since axial and lateral displacements are orthogonal to each other,they can be decoupled and evaluated separately, and then be combined together to determine the finaldisplacement. For example, the process of finding a 2× 2 template in a 4× 4 search window using DCC(Fig. 3(b)) is as follows: (i) γ(u, v) is calculated at each position in the center column to find γmax,y,(ii) γ(u, v) is calculated at each position in the row containing γmax,y to find γmax,x, and (iii) the (x, y)coordinate determined from the above two steps is regarded as the final position of the template. Thisscheme converts 2-D matrix calculation into 1-D array calculations, resulting in significant reduction ofcomputational load. If N ×N template is searched in a M ×M window, DCC needs 2× (M −N + 1)times of calculations, compared to (M −N + 1)2 for FNCC. Template size of 20× 20 pixels and searchwindow twice the size of template (40 × 40 pixels) are common, for which the numbers of calculationsper grid point is 42 for DCC, compared to 441 for conventional algorithm, achieving 90% reduction inthe computational load.

DIC adopting DCC algorithm was custom-developed using Matlab and the results were comparedwith those from FNCC, since FNCC has been widely accepted as a standard pattern matching algorithmfor DIC [20–22].

2.2. Sample preparation

2.2.1. Gelatin phantomsGelatin-based phantoms were designed to contain an inclusion with higher stiffness than the surround-

ing matrix, mimicking a carcinoma in a normal breast tissue [23]. Gelatin inclusions and gelatin matriceswere made with the same constituents to have the similar echogenicity, i.e., 1 wt% agarose (J.T. Baker),2 wt% glutaraldehyde (Sigma-Aldrich), 5 wt% n-propanol (Fisher-Scientific), gelatin (Fluka) (20 wt%for inclusion, 3 wt%, 5 wt% and 10 wt% for matrices, respectively), and distilled water (the remainingwt%). The procedure to fabricate the phantoms is described in detail in Ref. [24].

The fabricated phantoms had different stiffness ratio of the inclusion to the matrix, as the stiffnessof the matrix was varied with the change of gelatin content, while that of the inclusion was fixed. Thefabricated phantoms are called 3%, 5% and 10% gelatin phantoms, respectively, following the gelatincontent in the matrix.

2.2.2. Compression samplesTo measure the mechanical properties of the inclusion and the matrix, five cylindrical samples with

aspect ratio of 1 (height and diameter 4 cm each) were additionally made using a Teflon mold at eachformula.

2.2.3. Commercial breast phantomsCommercial breast phantom (Blue Phantom) containing a broad range of elastic masses was used to

generate elastograms (Fig. 4). These masses had different echotextures such as hypoechoic, isoechoicand hyperechoic.


Fig. 4. Ultrasound images of commercial breast phantom (Elastography Ultrasound Breast Phantom, Blue Phantom) wereobtained by conventional ultrasound system (Accuvix XQ, Medison). (Colors are visible in the online version of the article;http://dx.doi.org/10.3233/BME-140980.)

2.3. Compression test

Compression tests were conducted on compression samples with a TA material testing machine (Sta-ble Micro Systems) with a 50 N load cell. Each sample was loaded up to the engineering strain of 15%at three different loading speeds: 10 µm/s, 100 µm/s and 1000 µm/s. Five sets of tests were performed ateach formula using different samples and the results were averaged.

2.4. Stiffness ratio for nonlinear-elastic materials

Most biological tissues demonstrate nonlinear stress–strain relationships for sufficiently large defor-mations [25,26] and different nonlinear behaviors were demonstrated by different breast tissue typesincluding tumors. For example, Krouskop et al. [27] and Wellman [28] reported that cancers are notonly much stiffer than the surrounding tissue but also exhibit greater nonlinearity. Various models havebeen employed to express the nonlinear stress–strain behavior of biological tissues including hypere-lastic [29], exponential and power law [30]. Among them power law function in Eq. (4) is simple, butprovides reasonably accurate fittings [26,27]:

σ = Cεn. (4)

Stiffness of nonlinear elastic material can be represented by the tangent modulus at a specific strain [31].For Eq. (4),

Et =dCεn

dε=

n

εσ. (5)

When a phantom is deformed under compression, the stiffness ratio of the inclusion to the matrix can beestimated using force equilibrium along the center line of the inclusion:

Et,i

Et,m=

ni/εinm/εm

σiσm

=εmεi

ni

nm. (6)


Meanwhile, the exponent ratio (ni/nm) can be estimated by manipulating the strain values at differentdegrees of deformation. By taking the log of Eq. (4) for the inclusion and the matrix,

logσi = logCi + ni log εi,(7)

logσm = logCm + nm log εm.

Subtracting and rearranging Eq. (7) yields

log εm =1nm

log

(Ci

Cm

)+

ni

nmlog εi, (8)

which is the equation of a straight line in the form of Y = a + bX where Y = log εm and X = log εi.From the measured strains (X1,Yi), . . . , (Xn,Yn), a and b can be determined using least squares method:

a =

∑X2

i

∑Yi −

∑Xi

∑XiYi

n∑

X2i − (

∑Xi)2

, b =n∑

XiYi −∑

Xi∑

Yi

n∑

X2i − (

∑Xi)2

. (9)

Then the b can be plugged into Eq. (6) with the strain ratio (εm/εi) estimated on the elastograms todetermine the stiffness ratio of the lesion to the matrix.

2.5. Ultrasound imaging

Fabricated gelatin phantoms were uniaxially compressed to reach 15% strain at the loading rate of10 µm/s. During compression, US images were taken with the US probe (L6-12IS, 6–12 MHz) placedin the direction perpendicular to the axis of cylindrical inclusion to image its circular cross-section(Fig. 5(a)). For the US imaging of commercial breast phantom, the same US probe was applied to thephantom and moved slightly inferior and superior to obtain the images under compression (Fig. 4).

Human breast US images were taken by the ultrasound technician in the Grand River Hospital (Kitch-ener, Ontario) using Philips IU22 XMTRAIX US system with a broadband linear array transducer(L17-5, 5–17 MHz).

2.6. Displacement-gradient elastogram

Grid points were generated in undeformed US image, and the movements were tracked in the sub-sequent deformed images (Fig. 5(b)). In order to suppress the decorrelation errors in large deformation[10], dynamic referencing scheme [21] (also known as multicompression [31]) was employed, in whichmost recent image was used as a reference image and the summation of incremental displacements wastaken as the total displacement. Then the displacement-gradients at each grid point were taken as strainsas:

εx =∂ux∂x

, εy =∂uy∂y

. (10)

In order to reduce roughness of strain field caused by random noise and sub-pixel errors, smoothingalgorithm based on the penalized least square regression method was applied to the strain fields [24,32].Graphical representation of the strain fields generated by the above method is called DGE elastogram.


Fig. 5. Ultrasound images of a 5% gelatin phantom (20% gelatin inclusion in 5% gelatin matrix) containing a circular inclusion:(a) undeformed, and (b) deformed by 15% compression, and displacement fields determined by DIC in: (c) axial (x-) and (d) lat-eral directions, respectively, estimated by DCC algorithm. Grid points in undeformed image (green dots in (b)) were tracked byDIC to determine the new locations of grid points in deformed image (red dots in (b)). The arrow in the right-hand side of (b) in-dicates the loading direction. (The colors are visible in the online version of the article; http://dx.doi.org/10.3233/BME-140980.)

3. Results and discussion

3.1. Validation of DCC algorithm

Based on the movements of the grid points tracked by the DCC algorithm (Fig. 5(b)), displacementsat each grid point in both axial and lateral directions (Fig. 5(c) and (d), respectively) were estimated.The difference of the displacements estimated by FNCC and DCC were quantified using standard errorsin both directions:

Ex =

√√√√ 1n

n∑i=1

(uDCC,i − uFNCC,i)2, Ey =

√√√√ 1n

n∑i=1

(vDCC,i − vFNCC,i)2, (11)

where n is the number of grid points, and uFNCC,i, uDCC,i and vFNCC,i, vDCC,i are axial and lateral direc-tional displacements of the grid points estimated by FNCC and DCC, respectively.


Fig. 6. (a) Standard displacement errors in axial and lateral directions with respect to applied strain. Errors were determined bycomparing the displacements from FNCC and DCC algorithms; (b) Image processing times for FNCC and DCC algorithms.11 images with different size of templates around 13 × 13 grid points were processed. Search window size was twice of thetemplate size. (Colors are visible in the online version of the article; http://dx.doi.org/10.3233/BME-140980.)

Figure 6(a) showed that although the errors were increasing with the increase of applied strain, theywere almost negligible across all applied strains. Furthermore, most of the errors occurred at the outer-most grids which did not affect the feature of an elastogram.

Computational efficiency was assessed using the CPU time spent for image processing by DCC andFNCC algorithms. CPU time was tracked by “Profiler” tool in Matlab with only one CPU in Intel i7Quad core CPU (2.00 GHz) was set active.

Figure 6(b) clearly indicates that as the template size (and hence the search window size) increased,CPU time for FNCC processing increased very rapidly; while that for DCC was almost linear and muchless than FNCC across all template sizes, particularly for the template size larger than 20 × 20 pixels.

It is generally known that higher tracking accuracy can be achieved by employing larger templatesize [33]. To the author’s experience, template size of 20 × 20 pixels and the search window of twicethe template size provide reasonable balance between accuracy and computational efficiency. Since theDIC requires other functions such as data reading and writing, results printing and post processing,pure image processing time should be less than 0.5 s to achieve near quasi real-time elastogram at theframe rate of 1 fps. Since FNCC spent 1.5 s of CPU time to process each image, 1 fps could never beachieved unless a much faster, highest-end workstation is employed. On the other hand, DCC required


only 0.35 s CPU time per each image; thus quasi real-time display could be accomplished with medium-level computers (Intel i7 2.0 GHz CPU and 8 GB RAM).

From the above validation tests, it was concluded that the adoption of DCC algorithm to DIC methodcould save significant amount of computational load while preserving the pattern matching accuracy.All the elastograms presented in the rest of the paper were generated by the DIC incorporating DCCalgorithm.

3.2. Stress–strain curves

Representative engineering stress–strain curves from simple compression tests at the loading rate of10 µm/s are shown in Fig. 7. All of the curves demonstrate nonlinear stress–strain behaviors in the testedregion. Curve fittings to the loading branches of stress–strain curves were attempted using the power lawfunctions in Eq. (4) and the best fitting C and n values are presented in Table 1. Almost perfect fitting

Fig. 7. Engineering stress–strain curves of gelatin samples with various gelatin contents (20%, 10%, 5% and 3%) at the loadingrate of 10 µm/s. (Colors are visible in the online version of the article; http://dx.doi.org/10.3233/BME-140980.)

Table 1

The constants and the R-squared values of exponential functions fitted to the stress–strain curves in Fig. 5

C (kPa) n R2

20% gelatin 75 1.3 1.000010% gelatin 36 1.28 1.00005% gelatin 22 1.26 0.99993% gelatin 4.7 1.2 0.9998∗Normal fat 21 1.08 0.9913∗Glandular tissue 140 1.39 0.9954∗Ductal carcinoma in situ 5833 2.86 –∗Carcinoma 1746 2.20 –

Notes: For comparison purpose, the constants of the exponential functions fitted to breast tissue data in Ref. [27] (∗) arepresented. R-squared values for ductal carcinoma in situ (DCIS) and carcinoma, R-squared values are not calculated, sinceonly two data points are available.


could be achieved as indicated by R-squared values. The data for benign and malignant breast tissuesfrom Ref. [27] were fitted using the same type of functions, which also resulted in excellent agreements(Table 1). It is particularly notable that values of C and n for malignant tissues were much higher thanbenign tissues.

3.3. DGE elastograms

DGE elastograms were generated from the B-mode images of the 5% gelatin phantom compressedby 5%, 10% and 15% strains (Fig. 8). While the existence of stiff inclusion could be clearly identifiedon both axial and lateral elastograms at all strains, the shape and the size of the inclusion (dotted curve)were better illustrated on lateral elastograms (Fig. 8(b), (d) and (f)). On the other hand, axial elastogramscould provide the information on the strain ratio of the inclusion to the matrix (εi/εm). The strains at

Fig. 8. Axial and lateral elastograms of a 5% gelatin phantom compressed at: 5% strain ((a) and (b)), 10% strain ((c) and (d))and 15% strain ((e) and (f)). (Colors are visible in the online version of the article; http://dx.doi.org/10.3233/BME-140980.)


Table 2

Strains and stiffness ratios of 5% gelatin phantom

εm εi εm/εi ni/n∗m Et,i/Et,m

From Eq. (6) From the trend curve−0.025 −0.0112 2.232 1.0425 2.330 2.385−0.05 −0.02 2.381 (1.0317) 2.485 2.405−0.075 −0.031 2.419 2.496 2.433−0.098 −0.040 2.390 2.466 2.468−0.123 −0.050 2.412 2.488 2.484−0.148 −0.061 2.426 2.503 2.498∗ni/nm value was calculated by Eq. (6) using the strains on the elastograms. On the other hand, the value in ( ) was directlycalculated from the n values in Table 1.

the center of the inclusion (εi) and at the remote region in the matrix (εm) estimated from the axialelastograms (Fig. 8(a), (c) and (e)) are presented in Table 2.

3.4. Stiffness ratio in large deformation

By substituting εi and εm values in Table 2 into Eq. (9), ni/nm for this phantom was estimated to be1.0425 which was very close to 1.0317 directly calculated from the fitting exponents in Table 1. Usingthis value and the strain ratio (εm/εi), the stiffness ratio of the inclusion to the matrix (Et,i/Et,m) couldbe calculated using Eq. (6) (Table 2). Due to the nonlinearities in stress–strain curves, the stiffness ratiowas not a single value, but showed an increasing trend with the increase of deformation. Stiffness ratioswere also slightly higher than strain ratios, due to the ni/nm larger than unity.

For the comparison, Et,i and Et,m values for 20% and 5% gelatin, respectively, were directly calcu-lated from the trend curves (using Eqs (4) and (5)), and stiffness ratio (Et,i/Et,m) was calculated fromthem (Table 2). Good agreements between both sets of stiffness ratios in the magnitude and the trendsuggest that the variation of stiffness ratio for large strain region over 10% can be addressed by DGEelastograms. Currently only strain ratio and the morphology of the lesion are used for the diagnosis ofbreast cancer [12,34]. However, malignant tissues are known to demonstrate highly nonlinear behaviorwith the higher exponent values than benign tissues [27] (Table 1). Therefore, the difference betweenthe strain ratio and stiffness ratio, and the exponent ratio may be considered as potential indicators todifferentiate benign and cancerous lesions.

3.5. Effect of stiffness ratio on elastograms

In general, larger stiffness difference between the lesion and the surrounding tissue generates highercontrast on the elastogram. The ability of DGE in delineating the lesions with different stiffness wasinvestigated by varying the stiffness ratio of the inclusion to the matrix of gelatin phantom in two ways.

First, elastograms were generated using a 3% gelatin phantom (20% inclusion in 3% matrix). Theinclusion has an irregular shape, mimicking natural carcinoma in the breast (Fig. 9(a)). Since 3% gelatinis much softer than 5% one; the stiffness ratio of 3% gelatin phantom was much higher than that of5% phantom. As a result, the inclusion was significantly less deformed than the surrounding matrix, asindicated by the strain color bars in Fig. 9(c) and (e). It was particularly notable across all applied strainsthat the inclusion in axial elastogram appeared to be much smaller than its actual size. The shape wasalso significantly distorted, so that it was hard to identify the actual shape and the size from the axial


Fig. 9. US image of a 3% gelatin phantom (20% gelatin inclusion in 3% gelatin matrix) containing an irregular-shaped inclusionphantom: (a) undeformed, and (b) deformed by 15% compression. Axial and lateral elastograms at: 10% strain ((c) and (d))and 15% strain ((e) and (f)). (Colors are visible in the online version of the article; http://dx.doi.org/10.3233/BME-140980.)

elastogram. On the other hand, lateral elastograms successfully delineated the irregular-shaped inclusion(Fig. 9(d) and (f)), and the size and the shape could be accurately identified.

Second, DGE was performed on 10% gelatin phantom (20% inclusion in 10% matrix). Since thegelatin content in the matrix was twice that of the 5% phantom, the stiffness ratio of the inclusion to thematrix was only about 2 across the entire strain region (Table 3); nevertheless, Fig. 10 indicated that thedetection performance of DGE was not much degraded. When the applied strain was small (ε < 10%),the inclusion was barely recognizable on axial elastogram (Fig. 10(a)); however, it could still be clearlyidentified on lateral elastogram (Fig. 10(b)). As the applied strain increased, the inclusion became highlynoticeable on both elastograms (Fig. 10(c) and (d)). Also, the strains in the inclusion and the matrix couldbe estimated on axial elastograms with reasonable accuracy to evaluate the exponent ratio and stiffnessratios (Table 3).


Table 3

Strains and stiffness ratios of 10% gelatin phantom

εm εi εm/εi ni/n∗m Et,i/Et,m

From Eq. (6) From the trend curve0.025 0.015 1.56 1.0143 1.679 1.6900.048 0.03 1.60 (1.0156) 1.623 1.7290.075 0.046 1.63 1.654 1.7350.1 0.06 1.67 1.691 1.7340.125 0.075 1.67 1.691 1.7410.15 0.089 1.69 1.710 1.742∗ni/nm value was calculated by Eq. (6), while the value in ( ) was directly calculated from the n values in Table 1.

Fig. 10. Axial and lateral elastograms of a 10% gelatin phantom (20% gelatin inclusion in 10% gelatin matrix) at: 10%applied strain ((a) and (d)); 15% applied strain ((c) and (d)). (Colors are visible in the online version of the article;http://dx.doi.org/10.3233/BME-140980.)

3.6. Elastography phantom

DGE was applied to commercial breast phantom in the same manner as clinical practice. Breast phan-tom contains several masses, mimicking benign and malignant lesions, and Fig. 11(a) is the US imageof one of hyperechoic masses. Grid points were generated in the mass and the upper part of the imageto generate the elastograms (Fig. 11(b)). Figure 11(c) is the DGE elastogram when the phantom wascompressed to approximately 10% strain, which indicates that strain level within the mass was less thana tenth of the remote matrix region. The elastogram at 30% strain in Fig. 11(d) also presented similar be-havior. Note that even at such high strains, degradations in the elastogram images were minimal and the


Fig. 11. US images of a hyperechoic mass in a commercial phantom (Blue Phantom): (a) undeformed, and (b) deformed by30% compression; DGE elastograms at: 10% strain (c) and 30% strain (d); US image of the same region (e) and time-gradientelastogram (f) provided by phantom manufacturer (Blue Phantom). (Colors are visible in the online version of the article;http://dx.doi.org/10.3233/BME-140980.)

features of the strain field were maintained. Using the exponent ratio and strain ratios estimated from theelastograms, stiffness ratios could be successfully determined (Table 4). Both strains and stiffness ratiossuggested that the mass was approximately 10 times stiffer than the matrix, thus, could be malignant.

On the other hand, Fig. 11(e) and (f) are the US image and time-gradient elastogram generated byhigh-end US scanner (Philips xMATRIX) provided by the phantom manufacturer (Blue Phantom). The


Table 4

Stiffness ratios of hyperechoic mass to the matrix in the breast elastography phantom

εi εm ni/nm εm/εi Et,i/Et,m

−0.002 −0.05 0.6529 25 16.32−0.006 −0.10 20 13.05−0.010 −0.15 15 9.79−0.020 −0.20 13.33 8.70−0.025 −0.25 10.87 7.10−0.040 −0.30 10 6.53

elastogram shows highly noisy patterns and only qualitative measurements are available (SF: soft; HD:hard).

Figure 12(a) is the US image of another mass in the same phantom. This mass appears almost black,which indicates that it is an anechoic mass. The movements of grid points in the US image at around 20%compression (Fig. 12(b)) suggested that the applied deformation was not purely compressive, but unin-tended lateral movement was induced. Nonetheless, relatively uniform and smooth elastograms wereobtained by DGE at both 10% and 20% strains (Fig. 12(c) and (d), respectively). Note that the lesionwas hardly noticeable on both elastograms, which suggested that this lesion might have similar stiffnessas the surrounding tissues. Considering the anechoic property on US images, this mass was deemed to bea cystic lesion which was not malignant. The elastogram generated by commercial elastography system(Fig. 12(e)) also demonstrated similar behavior.

3.7. Human breast imaging

The feasibility of DGE for clinical application was examined by applying the DGE to the US imagesof human breasts containing suspicious lesions. Figure 13(a) is the image of the lesion in BI-RADS 4category and Fig. 13(b) is the DGE elastogram of the region of interest (ROI) in Fig. 13(a). The elas-togram suggests that this lesion is much stiffer (>5 times) than the surrounding tissue and the size of thelesion appears much larger than that in US image, which is frequently used as a parameter for differen-tiating benign and malignant lesions [12]. Therefore, this lesion was deemed to be a malignant tumor,which was confirmed by the subsequent biopsy as an invasive mammary carcinoma.

The DGE was also applied to BI-RADS 2 lesion in Fig. 12(c) to generate the elastogram of this lesionas shown in Fig. 13(d). The lesion looked anechoic on US image and slightly softer than surroundingtissue on elastogram, thus it might be a simple cyst, which was confirmed by cytology.

The above clinical study was only preliminary and the number of cases was too small to be statisticallyrelevant; however, it demonstrated the potential of DGE as a diagnostic tool for breast cancer detection.Further studies will be conducted to investigate various parameters as diagnostic estimators and to assessthe threshold of each estimator for the detection of malignant tumours.

4. Conclusions

The displacement-gradient elastography (DGE) was developed by applying digital image correla-tion method and smoothing algorithm to the B-mode US images from conventional ultrasound scan-ner. Near-real time elastogram display could be achieved by employing the newly developed pattern


Fig. 12. US images of an anechoic mass in a commercial phantom: (a) undeformed, and (b) deformed by 20% compression;DGE elastograms at: 10% strain (c) and 20% strain (d); US image of the same region and time-gradient elastogram (e) providedby phantom manufacturer (Blue Phantom). (Colors are visible in the online version of the article; http://dx.doi.org/10.3233/BME-140980.)


Fig. 13. (a) US image of a human breast containing BI-RADS 4 lesion, (b) the DGE elastogram of the ROI in (a), (c) US imageof BI-RADS 2 lesion, and (d) the DGE elastogram of the ROI in (c). (Colors are visible in the online version of the article;http://dx.doi.org/10.3233/BME-140980.)

matching algorithm. Performance of DGE was evaluated using gelatin and commercial breast phan-toms. It was verified that DGE could detect the inclusion with a wide range of stiffness with little imagedegradation. It was also proposed that lateral elastograms might be useful in estimating the size andshape of the lesion, while stiffness and exponent ratios could be evaluated by axial elastograms. Thecomparisons between DGE elastograms and conventional elastograms on the commercial breast phan-tom suggested that DGE might have comparable performance in diagnosing the pathological lesionto time-gradient elastography. The DGE elastograms of human breasts also showed the potential ofDGE for clinical applications, which needs to be further investigated through the study on more clinicalcases.

Acknowledgement

The work was supported by Natural Sciences and Engineering Research Council of Canada (NSERC).


References

[1] A. Jemal, R. Siegel, J. Xu and E. Ward, Cancer statistics, CA Cancer J. Clin. 60(5) (2010), 277–300.[2] C. DeSantis, R. Siegel and P. Bandi, Breast cancer statistics, CA Cancer J. Clin. 61(6) (2011), 408–418.[3] C. Di Maggio, State of the art of current modalities for the diagnosis of breast lesions, Eur. J. Nucl. Med. Mol. Imaging

31(1) (2004), S56–S69.[4] R.A. Smith, V. Cokkinides and H.J. Eyre, American Cancer Society guidelines for the early detection of cancer, CA

Cancer J. Clin. 56(1) (2006), 11–25.[5] C.H. Lee, D.D. Dershaw, D. Kopans, P. Evans, B. Monsees, D. Monticciolo et al., Breast cancer screening with imaging:

Recommendations from the Society of Breast Imaging and the ACR on the use of mammography, breast MRI, breastultrasound, and other technologies for the detection of clinically occult breast cancer, J. Am. Coll. Radiol. 7(1) (2010),18–27.

[6] L. Esserman, Y. Shieh and I. Thompson, Rethinking screening for breast cancer and prostate cancer, Amer. Med. Assoc.302(15) (2009), 1685–1692.

[7] P. Skrabanek, Mass mammography: The time for reappraisal, Int. J. Tech. Ass. Health Care 5(3) (1989), 423–430.[8] L. Liberman, L. Liberman, A.F. Abramson, F.B. Squires, J.R. Glassman, E.A. Morris et al., The breast imaging reporting

and data system: positive predictive value of mammographic features and final assessment categories, Am. J. Roentgenol.17(1) (1998), 35–40.

[9] K.J. Parker, M.M. Doyley and D.J. Rubens, Imaging the elastic properties of tissue: the 20 year perspective, Phys. Med.Biol. 56(1) (2011), R1–R29.

[10] J. Ophir, S.K. Alam, B.S. Garra, F. Kallel, E.E. Konofagou, T. Krouskop et al., Elastography: Imaging the elastic propertiesof soft tissues with ultrasound, J. Med. Ultrason. 29(Winter) (2002), 155–171.

[11] J. Ophir, S.K. Alam, B. Garra, F. Kallel, E. Konofagou, T. Krouskop et al., Elastography: ultrasonic estimation and imagingof the elastic properties of tissues, Proc. Inst. Mech. Eng. Part H, J. Eng. Med. 213(3) (1999), 203–233.

[12] J. Ophir, I. Cespedes, B. Garra, H. Ponnekanti, Y. Huang and N. Maklad, Elastography: Ultrasonic imaging of tissue strainand elastic modulus in vivo, Eur. J. Ultrasound. 3(1) (1996), 49–70.

[13] P.-L. Yen, D.-R. Chen, K.-T. Yeh and P.-Y. Chu, Development of a stiffness measurement accessory for ultrasound inbreast cancer diagnosis, Med. Eng. Phys. 33(9) (2011), 1108–1119.

[14] A.R. Skovoroda, A.N. Klishko, D.A. Gusakyan, Y.I. Mayevskii, V.D. Yermilova, G.A. Oranskaya et al., Quantitativeanalysis of the mechanical characteristics of pathologically changed soft biological tissues, Biophysics 40(6) (1995),1359–1364.

[15] T. Rago, F. Santini, M. Scutari, A. Pinchera and P. Vitti, Elastography: new developments in ultrasound for predictingmalignancy in thyroid nodules, J. Clin. Endocrinol. Metab. 92(8) (2007), 2917–2922.

[16] H. Zhi, B. Ou, B.-M. Luo, X. Feng, Y.-L. Wen and H.-Y. Yang, Comparison of ultrasound elastography, mammography,and sonography in the diagnosis of solid breast lesions, J. Ultrasound. Med. 26(6) (2007), 807–815.

[17] A. Itoh, E. Ueno, E. Tohno, H. Kamma, H. Takahashi, T. Shiina et al., Breast disease: Clinical application of us elastog-raphy for diagnosis, Radiol. 239(2) (2006), 341–350.

[18] J.M. Mari and C. Cachard, Acquire real-time RF digital ultrasound data from a commercial scanner, Electron. J. Tech.Acoust. 3 (2007), 1–16.

[19] T.C. Chu, W.F. Ranson, M.A. Sutton and W.H. Peters, Applications of digital-image-correlation techniques to experimen-tal mechanics, Exp. Mech. 25(3) (1985), 232–244.

[20] J.P. Lewis, Fast template matching. Vision Interface 95, Canadian Image Processing and Pattern Recognition Society,Quebec City, Canada, 1995, pp. 120–123.

[21] Y. Han, A.D. Rogalsky, B. Zhao and H.J. Kwon, The application of digital image techniques to determine the largestress–strain behaviors of soft materials, Polym. Eng. Sci. 52(4) (2011), 826–834.

[22] H.J. Kwon, A.D. Rogalsky, C. Kovalchick and G. Ravichandran, Application of digital image correlation method to biogel,Polym. Eng. Sci. 50(8) (2010), 17–20.

[23] E.L. Madsen, J.A. Zagzebski and G.R. Frank, An anthropomorphic ultrasound breast phantom containing intermediate-sized scatterers, Ultrasound. Med. Biol. 8(4) (1982), 381–392.

[24] Y. Han, D.W. Kim and H.J. Kwon, Application of digital image cross-correlation and smoothing function to the diagnosisof breast cancer, J. Mech. Behav. Biomed. 14(2012) (2012), 7–18.

[25] P.R. Fung, Biomechanics: Mechanical Properties of Living Tissues, 2nd edn, Springer-Verlag, New York, 1993.[26] Y.C.B. Fung, Elasticity of soft tissues in simple elongation, Am. J. Physiol. 213(6) (1967), 1532–1544.[27] T.A. Krouskop, T. Wheeler, F. Kallel and B. Garra, Elastic moduli of breast and prostate tissues under compression,

Ultrason. Imaging 20(4) (1998), 260–274.[28] P. Wellman, Tactile imaging, PhD thesis, Harvard University, 1999.[29] J.J. O’Hagan and A. Samani, Measurement of the hyperelastic properties of 44 pathological ex vivo breast tissue samples,

Phys. Med. Biol. 54(8) (2009), 2557–2569.


[30] A.L. Kellner and T.R. Nelson, Simulation of mechanical compression of breast tissue, IEEE Trans. Biomed. Eng. 54(10)(2007), 1885–1891.

[31] T.Z. Pavan, E.L. Madsen, G.R. Frank, J. Jiang, A.A.O. Carneiro and T.J. Hall, A nonlinear elasticity phantom containingspherical inclusions, Phys. Med. Biol. 57(15) (2012), 4787–4804.

[32] G. Wahba, Spline Models for Observational Data, CBMS-NSF Regional Conference Series in Applied Mathematics,Vol. 59, Society for Industrial and Applied Mathematics, Philadelphia, PA, 1990, pp. 45–65.

[33] E. Verhulp, B. van Rietbergen and R. Huiskes, A three-dimensional digital image correlation technique for strain mea-surements in microstructures, J. Biomech. 37(9) (2007), 1313–1320.

[34] A. Thomas, F. Degenhardt, A. Farrokh, S. Wojcinski, T. Slowinski and T. Fisher, Significant differentiation of focal breastlesions, Acad. Radiol. 17(5) (2010), 558–563.

Date post:	18-Mar-2020
Category:	Documents
Upload:	others
View:	3 times
Download:	0 times

Low-cost quasi-real-time elastography using B-mode ultrasound...

Documents