Computer-aided tumor diagnosis using shear wave breast elastography

Ultrasonics 78 (2017) 125–133

Contents lists available at ScienceDirect

Ultrasonics journal homepage: www.elsevier.com/locate/ultras

Computer-aided tumor diagnosis using shear wave breast elastography Woo Kyung Moon a, Yao-Sian Huang b, Yan-Wei Lee b, Shao-Chien Chang b, Chung-Ming Lo c, Min-Chun Yang b, Min Sun Bae a, Su Hyun Lee a, Jung Min Chang a, Chiun-Sheng Huang d, Yi-Ting Lin e, Ruey-Feng Chang b,e,f,⇑ a

Department of Radiology, Seoul National University Hospital and Seoul National University College of Medicine, Republic of Korea Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan Graduate Institute of Biomedical Informatics Taipei Medical University, Taipei, Taiwan d Department of Surgery, National Taiwan University Hospital, Taipei, Taiwan e Graduate Institute of Network and Multimedia, National Taiwan University, Taipei, Taiwan f Graduate Institute of Biomedical Electronics and Bioinformatics, National Taiwan University, Taipei, Taiwan b c

a r t i c l e

i n f o

Article history: Received 18 July 2016 Received in revised form 16 January 2017 Accepted 13 March 2017 Available online 15 March 2017 Keywords: Elastography Shear wave Breast Tumor segmentation Computer-aided diagnosis

a b s t r a c t The shear wave elastography (SWE) uses the acoustic radiation force to measure the stiffness of tissues and is less operator dependent in data acquisition compared to strain elastography. However, the reproducibility of the result is still interpreter dependent. The purpose of this study is to develop a computeraided diagnosis (CAD) method to differentiate benign from malignant breast tumors using SWE images. After applying the level set method to automatically segment the tumor contour and hue-saturationvalue color transformation, SWE features including average tissue elasticity, sectional stiffness ratio, and normalized minimum distance for grouped stiffer pixels are calculated. Finally, the performance of CAD based on SWE features are compared with those based on B-mode ultrasound (morphologic and textural) features, and a combination of both feature sets to differentiate benign from malignant tumors. In this study, we use 109 biopsy-proved breast tumors composed of 57 benign and 52 malignant cases. The experimental results show that the sensitivity, specificity, accuracy and the area under the receiver operating characteristic ROC curve (Az value) of CAD are 86.5%, 93.0%, 89.9%, and 0.905 for SWE features whereas they are 86.5%, 80.7%, 83.5% and 0.893 for B-mode features and 90.4%, 94.7%, 92.3% and 0.961 for the combined features. The Az value of combined feature set is significantly higher compared to the B-mode and SWE feature sets (p = 0.0296 and p = 0.0204, respectively). Our results suggest that the CAD based on SWE features has the potential to improve the performance of classifying breast tumors with US. Ó 2017 Elsevier B.V. All rights reserved.

1. Introduction In recent years, a variety of manufacturers have begun to incorporate elastography, a real-time tissue stiffness measuring technique, in ultrasound equipment. The results show that the shear-wave elastography (SWE) has the ability to yield accurate differentiation of benign from malignant breast lesions [1–4]. The SWE uses the acoustic radiation force instead of the manual tissue compression and is less operator dependent in data acquisition compared to strain elastography [5,6]. The SWE is a technique based on the local estimation of shear-wave propagation speed. The elastic modulus (E) of linearly isotropic tissues is proportional ⇑ Corresponding author at: Department of Computer Science and Information Engineering, National Taiwan University, Taipei 10617, Taiwan. E-mail address: [email protected] (R.-F. Chang). http://dx.doi.org/10.1016/j.ultras.2017.03.010 0041-624X/Ó 2017 Elsevier B.V. All rights reserved.

to the square of shear-wave speed (E = 3qcs2; q, tissue density; cs, shear-wave velocity). From the velocity information, the elastic modulus (in kilopascal unit, [kPa]) can be calculated and displayed in a color-coded image with the option of measuring quantitative elasticity values. Several studies have reported that malignant breast masses show higher elasticity values than benign lesions on SWE images [4,7,8]. Using quantitative SWE features, both of the sensitivity and specificity of B-mode US can be improved [4]. However, the reproducibility of the result is still interpreter dependent and significant interobserver variability has been found in reader studies [9,10]. Therefore, a computer-aided analysis of SWE images is needed to evaluate lesion stiffness objectively and to aid in the task of classification of benign and malignant breast tumors. The computer-aided diagnosis (CAD) of elastography has been explored with SWE images to differentiate benign from malignant

126

W.K. Moon et al. / Ultrasonics 78 (2017) 125–133

breast tumors [11,12]. On histopathologic analysis, malignant breast tumors often accompanied by the dense fibrous tissues in the peritumoral area compared to benign breast tumors [13,14]. This finding is concordant with SWE results that the most common site of tumor-associated stiffness is generally in the surrounding stroma rather than the tumor itself [2,15]. In previous study [12], ten quantitative SWE features of the tumor and peritumoral areas, respectively (elasticity modulus mean, maximum and standard deviation, hardness degree and elasticity ratio) were proposed for diagnosing breast tumors. In this study, a CAD method is proposed based on SWE features extracted from surrounding tumor tissues for classification of benign and malignant breast tumors. The elasticity values of the tumor including the distribution of tissue strains and the distances of the stiff pixels are quantitated and analyzed to the tumor in the elastogram. The performance of a developed CAD based on SWE features is evaluated with a large database. In addition, B-mode features [16] are extracted and combined with the newly developed SWE features to diagnose breast tumors. 2. Materials

sured at the B-mode image was 4–9 mm in 34 lesions, 10–14 mm in 26 lesions, 15–19 mm in 15 lesions, 20–24 mm in 16 lesions, 25–29 mm in 3 lesions and 30–80 mm in 15 lesions. Furthermore, the size of lesions were 8–80 mm (mean, 26.62 ± 16.44 mm) for IDC, 8-50 mm (mean, 20.33 ± 13.82 mm) for DCIS, 15–33 mm (mean, 21.67 ± 8.06 mm) for ILC, 4–24 mm (mean, 10.85 ± 4.65 mm) for FA, 4–22 mm (mean, 9.14 ± 4.49 mm) for FCC, and 6–7 mm (mean, 6.5 ± 0.50 mm) for papilloma. The tumor size at the B-mode image is the largest diameter in the two orthogonal imaging planes used to record the three-dimensional tumor size. 2.2. US data acquisition In this study, the SWE was performed by experienced radiologists of breast US using the Aixplorer ultrasound system (SuperSonic Image, Les Jardins de la Duranne, Aix en Provence, France) with a 4–15 MHz linear transducer, SL15-4. The standard SWE image captured using the Aixplorer ultrasound system is illustrated in Fig. 1. Each SWE image contains the B-mode (bottom) and corresponding elastographic image of the target tumor and its surrounding tissues.

2.1. Lesions 3. The proposed method This retrospective study was approved by the local ethics committee and the informed consent was waived. The SWE images of 109 breast tumors (57 benign and 52 malignant) in 106 consecutive women (age range from 23 to 75 years, mean: 48.43 years, standard deviation: 9.85 years) were obtained prior to needle biopsy or surgical excision. There were 20 cases of fibroadenoma (FA), 35 cases of fibrocystic change (FCC) and 2 cases of papilloma in benign lesions. On the other hand, there were 42 cases of invasive ductal carcinoma (IDC), 6 cases of ductal carcinoma in situ (DCIS), 3 cases of invasive lobular carcinoma (ILC), and 1 metaplastic breast carcinoma in the malignant lesions. The tumor size mea-

In the proposed diagnostic scheme, the tumor contour needs to be delineated first for the purpose of extracting B-mode and SWE features for diagnosis. The tumor segmentation method is applied on the B-mode image to delineate the tumor contour so as to compute the B-mode features according to the tumor contour and gray level information. The segmentation procedure follows the same steps proposed in our previous study [17]. We compared manual segmentation to the automated scheme and the result of our automated scheme was excellent. A series of pre-processing steps are conducted on the B-mode image for the best tumor segmentation.

Fig. 1. The SWE images (a) A malignant case: 18 mm, invasive ductal carcinoma (ATE: 48.65, SS1: 17.04, SS2: 25.97, SS3: 4.40, SS4: 0.21, SS5: 0.05, NMD36: 0.02, NMD72: 0.02, NMD108: 0.85, NMD144: 1.00). (b) A benign case: 10 mm, fibrocystic change (ATE: 24.97, SS1: 24.08, SS2: 0.40, SS3: 0.00, SS4: 0.00, SS5: 0.00, NMD36: 1.00, NMD72: 0.00, NMD108: 0.00, NMD144: 0.00).

W.K. Moon et al. / Ultrasonics 78 (2017) 125–133

At first, to increase the gray-level difference between tumor and background regions, the sigmoid filter [18] is utilized to enhance the contrast of the original B-mode image. Next, the gradient magnitude filter [19] is applied to obtain the gradient image, which assist for extracting the edge information. Furthermore, to obtain a better segmentation result, the sigmoid filter is applied again to enhance the contrast of the gradient image. Then, the tumor contour is segmented by using the level set method [20,21] from the enhanced B-mode image and mapped on the corresponding SWE image. Fig. 2 shows a progression of images for the segmentation of the B-mode images. However, the proposed segmentation method applied on some lesions which connect with shadowing may result in the over segmentation. Hence, the controllable point is utilized for correcting the segmentation mask while the over segmentation happens [22]. For evaluating the possibility of a lesion to be malignant, both B-mode and SWE features are extracted for diagnosis. The Bmode features evaluate the morphology and texture information of the tumor. The eight B-mode feature classes are shape, orienta-

127

tion, margin, lesion boundary, echo pattern, posterior acoustic, speckle, and gray level co-occurrence matrix (GLCM) texture feature. The first three classes depict the geometric characteristics of the segmented tumor and the other classes depict its corresponding texture information. The details are described in our previous studies [16,23,24]. The SWE feature measures are taken from values of valid shear wave speed estimates of the tumor and peritumoral area in the SWE image. Originally, the tissue elasticity information is converted into a red-green-blue (RGB) color SWE image and superimposed on the B-mode image for easily recognizing the relationship between tissue strains and the lesion itself in the SWE. For recovering the shear elasticity data, the hue-saturation-value (HSV) color space transformation is applied to the RGB-coded SWE image [25]. Then, the SWE features are calculated by utilizing the elasticity information corresponding to the tumor region found in the Bmode image. Our previous study [26] had mentioned that the elasticity information is represented on the hue channel of the strain image. Compared to the research [12], the pre-process of subtract-

Fig. 2. The progression of images for the segmentation of the B-mode images (a) The original B-mode image. The image after applying (b) the sigmoid filter, (c) the gradient magnitude filter, (d) the sigmoid filter, (e) the contour after level set method. The contour superimposed on the original image (f).

128

W.K. Moon et al. / Ultrasonics 78 (2017) 125–133

mate overlapped on the B-mode image. Therefore, we use each pixel with valid shear wave speed estimate to develop a feature denoting overall tissue elasticity. The average tissue elasticity is the average kPa values within the tumor contour in the SWE image:

ATE ¼

1 X EðPÞ NR P2R

ð1Þ

where R is the region with tissue elasticities, NR is the number of pixels inside the tumor in R, and E(P) denotes the kPa value of pixel P. 3.2. Sectional stiffness ratio (SSR) Fig. 3. (a) The kPa scale of color bar ranging from 22 to 158, (b) the hue scale ranging from 0 to 360°.

ing original B-mode image from the superimposed image is needless to perform because of the properties of hue channel transformation. As illustrated in Fig. 3, the tissue elasticity in the color bar of SWE image is defined from 0 to 180 in RGB color space and the color of 0-degree (0 kPa) is denoted as blue whereas the range of hue value is from 0 to 360 and the color of 0-degree is red. Hence, for obtaining the elasticity value measured in kPa, the color bar is transformed into HSV color space. Then, the kPa is computed by mapping hue value of SWE image onto that of color bar directly. Unfortunately, we discover that the range of kPa value of color bar from 0 to 22 has the same hue value and the kPa value in the range cannot be obtained by mapping method. The same situation is also found in another range from 158 to 180. However, according to the literature [27], the medium kPa value of benign tumor is usually less than 80 kPa while that of malignant is larger than 100 kPa. Therefore, we define that the kPa values are 22 or 158 when the hue value falls in ranges from 0 to 22 or from 158 to 180 respectively for solving the problem of obtaining kPa value in these two ranges. This definition has no effect on classification results between benign and malignant tumor. Fig. 4 shows an example of the transformation from the SWE image to the hue image. After applying the modified HSV color transformation, the SWE features composed of three classes (average tissue elasticity, sectional stiffness ratio, and normalized minimum distance for grouped stiffer pixels) are calculated. 3.1. Average tissue elasticity (ATE) It’s worth noting that not all of the pixels in SWE image have valid elasticity values because the shear wave cannot propagate in liquid such as the breast cysts or some regions with depths larger than 3 cm [7,28]. That is, some pixels do not have valid shear wave speed esti-

The main concept of sectional stiffness ratio is derived from the stiffness degree which has been proven to be a representative feature to diagnose a tumor [17]. Ratios of pixel numbers in different elastic ranges to the total number of pixels with valid shear wave speed estimates are proposed as diagnostic features. The color bar representing tissue elasticity is divided into five elastic sections: Section 1 (22–36), Section 2 (37–72), Section 3 (73–108), Section 4 (109–144), and Section 5 (145–158). The five color sections and values were chosen since this have been most helpful to classify benign and malignant breast tumors in previous SWE studies and now used in clinical practice as a standard setting [2]. Hence, the feature, sectional stiffness ratios (SSR), is calculated within the region of tissue elasticity that is the range of white square in SWE image. The SSRn, n = 1. . .5, represent the ratio of pixels in the nth stiffness section to the total pixels with valid tissue elasticities:

SSRn ¼

EVSn NR

ð2Þ

where EVSn is the number of effective pixels whose elasticity values are in the nth elastic section, and NR is the total number of pixels with valid tissue elasticities. 3.3. Normalized minimum distance of grouped stiffer pixels (NMD) The distance of the nearest stiff pixels to the tumor is used as a diagnostic feature. At first, a stiffness threshold is needed to define the stiffer group of pixels. In the sectional stiffness ratio features, we define four threshold values (T) including 36, 72, 108, and 144 (those numbers were come from the kPa scale, ranging was 22–158, we quintile divide the kPa scale). This four threshold values are used to define four groups (G36, G72, G108, and G144) with different stiffness. Moreover, in order to limit the number of stiffer pixels in a SWE image, we found that top 1/16 of total elasticity

Fig. 4. An example of the transformation from the SWE image to the hue image (a) the original SWE image (b) the H channel image.

W.K. Moon et al. / Ultrasonics 78 (2017) 125–133

pixels have enough information in region of tissue elasticity by the experiments. Let GT’ be the group of stiffer pixels in GT ranked at top 1/16. The minimum distance MDT for group GT’ to the tumor boundary is defined as

MDT ¼ minP2G0T dis ðPÞ; T ¼ 36; 72; 108; and 144

ð3Þ

where dist(P) denotes the distance from pixel P to the tumor contour. Then each of the minimum distance MDT could be normalized by the maximum of four minimum distances:

NMDT ¼

MDT maxT¼36;

ð4Þ

72; 108; 144 ðMDT Þ

Table 1 listed the B-mode features that we calculated in our experiments. The B-mode features include eight categories: shape, orientation, margin, lesion boundary, echo pattern, posterior acoustic feature, speckle, and GLCM. Noting that when extracting the speckle feature [16], the equation is utilized to convert the gray value raw data:

Iðx; yÞ ¼ 10Gðx;yÞ=G0

ð5Þ

where G(x,y) is the pixel value in B-mode images and G0, which is a linearization factor related to the frequency of the transducer (G0 = 50 in this study), converts G(x,y) to a linear scale. After the conversion, I(x,y) is obtained as the acoustic intensity. Then, a moving 5  5 window is applied to find a region with a range of ratios of mean intensity to a standard derivation from 0.8 to 1.2 in the raw intensity image. If the region satisfies the condition, the center pixel

129

of the region is defined as speckle. Then, the speckle features could be found for the speckle pixels. 3.4. Statistical analysis In order to classify benign and malignant tumors, the linear support vector machine (SVM) [29] is applied on the acquired database to generate a probability between 0 and 1. The threshold applied on the predicted values is set to 0.5 to differentiate suspected tumors. In other words, a tumor is regarded as benign when the predicted value is less than the threshold, otherwise malignant. Furthermore, the 10-fold cross-validation method [30] is applied to estimate the generalization error of the SVM model. In each iteration, N/10 case are left out of all N cases and used to test the trained result with the remaining cases. A total of N/10 iterations of training and testing are executed. The diagnostic performance of Bmode features, SWE features, and combined features are evaluated and compared by statistical methods. The receiver operating characteristic (ROC) curve is plotted by using the ROCKIT software (C. Metz; University of Chicago, Chicago, IL, USA). The diagnosis performance is assessed according to six indicators: sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV) and the area under the ROC curve [31]. The chi-square test is applied on sensitivity, specificity, accuracy, PPV and NPV, and the z-test is applied on Az (the area under the ROC curve) to measure the performance differences between different feature sets. The p-value of each pairwise comparison less than 0.05 is regarded as statistically sig-

Table 1 Computerized B-mode features. Category

Feature

Definition

Shape

Sa Sb Ae Pe Sa/b SPR Ec OE Mu

The length of the semi-major axis of the best-fit ellipse The length of the semi-minor axis of the best-fit ellipse The area of the best-fit ellipse The perimeter of the best-fit ellipse Lengths of semi-major axis and semi-minor axes The ratio between perimeters of the tumor contour and the best-fit ellipse The area ratio between the intersection and the union of the tumor and the best-fit ellipse The included angle h between the major axis of the best-fit ellipse and the x-axis The number of lobular areas whose interior maximum distance is larger than 5 in the distance map generated according to the segmented tumor boundary The number of angular areas generated according to the distance map and the maximum inscribed circle of the segmented tumor boundary Mua = Mu + Ma The contrast information, i.e. average intensity difference, between the inner and outer bands around the tumor boundary The normalized contrast difference between all tumor pixels and the top 25% brightest pixels The difference of average gray intensity between top 25% brightest pixels within the tumor and all tumor pixels The difference of average gray intensity between the tumor and its 10-pixel width outer band The contrast information between the tumor and the region under the tumor

Orientation Margin

Ma

Lesion boundary Echo pattern

Posterior acoustic feature Speckle

Mua LB EPC EBd EOd PS Savgnum_i Smean_i

SSD_i Sgmean_i SgSD_i GLCM

Energy Entropy Correlation Inverse difference moment Inertia Cluster shade Cluster prominence Haralick’s correlation

The average number of speckle pixels in the tumor ROI i = 0, 5, 10, indicates band widths outward from the tumor boundary to form different tumor ROIs The average of the qualified grayscale mean/SD (computed in a 5  5 moving window) of each speckle pixel in the tumor ROI i = 0, 5, 10 The standard deviation of the qualified grayscale mean/SD (computed in a 5  5 moving window) of each speckle pixel i = 0, 5, 10 The average gray value of speckle pixels in the tumor ROI i = 0, 5, 10 The standard deviation of grayscales of speckle pixels in the tumor ROI i = 0, 5, 10 8 GLCM texture features Means and standard deviations of GLCM features computed in the entire tumor region are respectively named mGtn, sGtn (n = 1, 2,. . ., 8) Similarly, means and standard deviations of GLCM features computed in the speckled tumor region are respectively named mGsn, sGsn (n = 1, 2,. . ., 8)

130

W.K. Moon et al. / Ultrasonics 78 (2017) 125–133

nificant. All statistical analyses are computed using the SPSS software (SPSS, version 19 for Windows; SPSS, Chicago, IL, USA). The proposed system is implemented by the programming language C++ under the Microsoft Visual C++ 2013 (Microsoft, Redmond, WA, USA), operating with the Microsoft Windows 10 operating system (Microsoft, Redmond, WA, USA), and running on the Intel processor (3.50 GHz Quad-core machine with 16.00 GB RAM).

Table 3 Tests of performance differences between different feature sets. Each number is the computed p-value.

Accuracy Sensitivity Specificity PPV NPV Az a

4. Experiment results The backward feature elimination method [24] is applied to eliminate useless features from the proposed feature sets. The performance indicators of diagnosing on B-mode features, SWE features, and combined features are listed in Table 2. The sensitivity, specificity, accuracy, PPV and NPV of B-mode features are 86.5%, 80.7%, 83.5%, 80.4% and 86.8%, respectively; the sensitivity, specificity, accuracy, PPV and NPV of SWE features are 86.5%, 93.0%, 89.9%, 91.8% and 88.3%, respectively. The sensitivity, specificity, accuracy, PPV and NPV of combined features are 90.4%, 94.7%, 92.3%, 94.0% and 91.5%, respectively. The combined feature set has the best performance among all feature sets. Fig. 5 illustrates the ROC curves corresponding to the three feature sets. The Az value of B-mode feature set is 0.893, the Az value of SWE feature set is 0.905, and the Az value of combined feature set is

B-mode v.s. SWE

Combined v.s. SWE

Combined v.s. B-mode

0.163 1.000 0.052 0.094 0.804 0.7524

0.471 0.539 0.696 0.675 0.563 0.0204a

0.037a 0.539 0.022a 0.038a 0.419 0.0296a

The p-value is statistically significant different.

0.961. According to the Az values, the combined features set is also higher than the other two feature sets. Hence, the combined features set is regard as the best choice of tumor diagnosis. Furthermore, the Az of combined feature sets is significantly higher compared to B-mode and SWE feature sets (p = 0.0296 and p = 0.0204). The chi-square test is applied on sensitivity, specificity, accuracy, PPV and NPV, and the z-test is applied on Az to measure the performance differences between different feature sets, as listed in Table 3. Figs. 6 and 7 show a benign and a malignant cases correctly classified using the proposed SWE or combined feature sets but misclassified using the B-mode feature set.

4. Discussion Table 2 Performance indicators of diagnosing on B-mode, SWE, and combined features. Features

Sensitivity

Specificity

Accuracy

PPV

NPV

SWE

86.5% (45/52) 86.5% (45/52) 90.4% (47/52)

93.0% (53/57) 80.7% (46/57) 94.7% (54/57)

89.9% (98/109) 83.5% (91/109) 92.3% (101/109)

91.8% (45/49) 80.4% (45/56) 94.0% (47/50)

88.3% (53/60) 86.8% (46/53) 91.5% (54/59)

B-mode Combined

Note: Sensitivity = TP/(TP + FN), Specificity = TN/(TN + FP), Accuracy = (TP + TN)/(TP + TN + FP + FN), PPV = TP/(TP + FP), NPV = TN/(TN + FN). The number in parentheses represents the raw data in this experiment.

Fig. 5. The ROC curves of the B-mode, SWE and combined feature sets.

Previous clinical studies on the SWE [32,33] showed that the average or the maximum elasticity values and elasticity ratio defined as the elasticity of the lesion to the surrounding parenchyma of the tumor were useful to classify the tumors. In this study, the SWE features are divided into three classes: average tissue elasticity, sectional stiffness ratio, and normalized minimum distance of grouped stiffer pixels. The average tissue elasticity feature is calculated the average elasticity values of all elasticity pixels, which express the average stiffness of tumor surrounding areas. The sectional stiffness feature separates the color bar to several sections as recommended by Berg et al. [2] and calculated the ratio of the section on the total number of tissue elasticity. The normalized minimum distance feature indicates the distance between surrounding stiff tissue to the tumor. The experimental results that the median values of SWE features selected as the most useful features by the backward feature elimination method are shown in Table 4. In SWE, understanding the characteristics of shear wave propagation to various tissues and artifacts in the elastogram are also important to make correct interpretation. The shear wave can propagate in hard regions and deep regions but the SNR of the motion may be too low to get a valid shear wave speed estimate [3]. The shear wave in these lesions has substantial noise or minimal tissue displacement and is not accurately interpretable. In our study, the new proposed features including the sectional stiffness ratio and the normalized minimum distance of grouped stiffer pixels are based on computing the number of SWE pixels with valid shear wave speed estimate rather than the absolute elasticity values. Namely, calculating the number of SWE pixels can reduce the error caused by artifact pixels. Hence, our SWE features are less likely causing misclassifications comparing with the previous studies. Xiao at al. [12] proposed 10 quantitative elastographic features of the tumor and peritumoral areas and a support vector machine classifier was used for the classification of benign and malignant tumors. The Az value for their CAD system using the combination of elastographic features was significantly higher than the Az value for visual assessment by the radiologists using BI-RADS (0.97 vs.

131

W.K. Moon et al. / Ultrasonics 78 (2017) 125–133

Fig. 6. A benign case of 17 mm fibrocystic change correctly classified by the proposed combined features. (a) The original B-mode image. (b) The SWE image. (c) The tumor contour overlapped on the B-mode image. (d) The tumor contour overlapped on the SWE image. (ATE: 28.41, SS1: 21.45, SS2: 2.23, SS3: 0.00, SS4: 0.00, SS5: 0.00, NMD36: 1.00, NMD72: 0.00, NMD108: 0.00, NMD144: 0.00).

Fig. 7. A malignant case of 20 mm invasive ductal carcinoma correctly classified by the proposed combined features. But the diagnostic results made according on B-mode and SWE features are benign and malignant, respectively. (a) The original B-mode image. (b) The SWE image. (c) The tumor contour overlapped on the B-mode image. (d) The tumor contour overlapped on the SWE image. (ATE: 52.76, SS1: 10.05, SS2: 11.76, SS3: 2.89, SS4: 1.69, SS5: 0.06, NMD36: 0.90, NMD72: 0.90, NMD108: 0.90, NMD144: 1.00).

Table 4 The median values of selected as the most useful SWE features by the backward feature elimination method. Median

Benign Malignant

ATE

SS1

SS2

SS3

SS4

SS5

NMD36

NMD72

NMD104

NMD144

24.3 49.4

23.1 14.5

0.28 7.73

0 3.96

0 0.61

0 0.06

1 0.91

0 0.87

0 0.98

0 1

132

W.K. Moon et al. / Ultrasonics 78 (2017) 125–133

0.91, p < 0.05). In our study, more novel elastographic features such as sectional stiffness ratio and normalized minimum distance of grouped stiffer pixels were proposed. The results indicated that the tumor with higher elastographic features will be regard as malignant. Also, not only the elastographic features but also the B-mode US features were used in their study. In this paper, the B-mode (morphologic and textural) features were included in the diagnosis and the diagnosis performance could be further improved. Our results agree with the results of previous clinical studies that have demonstrated that SWE or a combination of SWE and conventional US findings can classify breast tumors better than the use of conventional US alone [2,6,14,32]. The tumors are correctly classified by combined features and B-mode features, but the diagnostic results made according to SWE features were sometimes incorrect. There are two reasons of misdiagnosing using SWE features. First, some regions around of the fascia or fibrosis in the lesion have high stiffness [33]. Second, the necrosis and heterogeneity in malignant tumor are cause of the tumor softening [10,33,34]. In some cases, the tumors are correctly classified by the combined features and SWE features, but the diagnostic results made according to B-mode features are incorrect. The misclassifications may be due to the shape of tumors. In the misclassified cases, the shape of benign tumor is irregular and the malignant tumor was regular. In some cases, the tumor was incorrectly classified by all feature sets. One reason may be related to the softening of malignant tumor mentioned above. The other reason might be the elasticity image missed some color signal or artifacts occurred, that lead to the tumor diagnosis was incorrectly classified. Continued research into understanding what we are measuring within the tissues is needed to fully understand and optimize SWE for breast lesion characterization [35,36]. Besides extracting the features to diagnose tumors, automatically segmenting the tumor contour is also the aim of this study. In order to extract the breast tumor contour, the level set segmentation method is used on the B-mode image. The method avoids the operator manually delineating the tumor contours in both the B-mode image and the elastogram so as to reduce the interobserver variability.

6. Conclusion and future works We developed a CAD method to differentiate benign from malignant breast tumors using SWE images. Sectional stiffness ratio and normalized minimum distance of grouped stiffer pixels as well as average tissue elasticity were the most important SWE features. The Az value of combined SWE and B-mode feature sets is statistically significant higher compared to B-mode and SWE feature sets (p = 0.0296 and p = 0.0204). Our results suggest that combining B-mode and SWE features has the potential to differentiate benign from malignant tumors. In future, the proposed segmentation method would be modified to satisfy the lesions with shadowing for reducing user dependence. Moreover, further studies are needed to evaluate the potential of our CAD system to improve physician performance to distinguish benign from malignant breast tumors and to reduce observer variability.

Acknowledgement The authors thank the Ministry of Science and Technology (MOST 359 104-2221-E-002-062-MY3) and National Taiwan University (NTU-ERP-105R890861) of the Republic of China for the financial support. This work was also supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIP) (2015R1A2A1A05001860).

References [1] J.H. Youk, E.J. Son, A.Y. Park, J.A. Kim, Shear-wave elastography for breast masses: local shear wave speed (m/sec) versus Young modulus (kPa), Ultrasonography 33 (2014) 34–39. [2] W.A. Berg, D.O. Cosgrove, C.J. Dore, F.K. Schafer, W.E. Svensson, R.J. Hooley, et al., Shear-wave elastography improves the specificity of breast US: the BE1 multinational study of 939 masses, Radiology 262 (2012) 435–449. [3] R.G. Barr, Sonographic breast elastography: a primer, J. Ultrasound Med. 31 (2012) 773–783. [4] P. Whelehan, A. Evans, K. Thomson, D. McLean, K. Brauer, C. Purdie, L. Jordan, L. Baker, A. Thompson, Quantitative shear wave ultrasound elastography: initial experience in solid breast masses, Breast Cancer Res. R 104 (2010). [5] R.G. Barr, Z. Zhang, Effects of precompression on elasticity imaging of the breast development of a clinically useful semiquantitative method of precompression assessment, J. Ultrasound Med. 31 (2012) 895–902. [6] A. Athanasiou, A. Tardivon, M. Tanter, B. Sigal-Zafrani, J. Bercoff, T. Deffieux, et al., Breast lesions: quantitative elastography with supersonic shear imaging–preliminary results, Radiology 256 (2010) 297–303. [7] J.M. Chang, W.K. Moon, N. Cho, A. Yi, H.R. Koo, W. Han, et al., Clinical application of shear wave elastography (SWE) in the diagnosis of benign and malignant breast diseases, Breast Cancer Res. Treat. 129 (2011) 89–97. [8] A. Tardivon, A. Athanasiou, M. Tanter, B. Sigal-Zafrani, J. Bercoff, T. Deffieux, J.L. Gennisson, M. Fink, S. Neuenschwander, Breast lesions: quantitative elastography with supersonic shear imaging—preliminary results, Radiology 256 (2010) 297–303. [9] S.H. Lee, J.M. Chang, W.H. Kim, M.S. Bae, M. Seo, H.R. Koo, et al., Added value of shear-wave elastography for evaluation of breast masses detected with screening US imaging, Radiology 273 (2014) 61–69. [10] S.H. Lee, N. Cho, J.M. Chang, H.R. Koo, J.Y. Kim, W.H. Kim, et al., Two-view versus single-view shear-wave elastography: comparison of observer performance in differentiating benign from malignant breast masses, Radiology 270 (2014) 344–353. [11] C.M. Lo, Y.C. Lai, Y.H. Chou, R.F. Chang, Quantitative breast lesion classification based on multichannel distributions in shear-wave imaging, Comput. Methods Programs Biomed. 122 (Dec 2015) 354–361. [12] Y. Xiao, J. Zeng, L. Niu, Q. Zeng, T. Wu, C. Wang, et al., Computer-aided diagnosis based on quantitative elastographic features with supersonic shear wave imaging, Ultrasound Med. Biol. 40 (2014) 275–286. [13] T.A. Krouskop, T.M. Wheeler, F. Kallel, B.S. Garra, T. Hall, Elastic moduli of breast and prostate tissues under compression, Ultrasonic Imaging 20 (1998) 260–274. [14] S.H. Lee, W.K. Moon, N. Cho, J.M. Chang, H.-G. Moon, W. Han, et al., Shear-wave elastographic features of breast cancers: comparison with mechanical elasticity and histopathologic characteristics, Investigative Radiol. 49 (2014) 147–155. [15] J. Zhou, W. Zhan, C. Chang, X. Zhang, Y. Jia, Y. Dong, et al., Breast lesions: evaluation with shear wave elastography, with special emphasis on the ‘‘stiff rim” sign, Radiology 272 (2014) 63–72. [16] W.K. Moon, C.M. Lo, C.S. Huang, J.H. Chen, R.F. Chang, Computer-aided diagnosis based on speckle patterns in ultrasound images, Ultrasound Med. Biol. 38 (2012) 1251–1261. [17] W.K. Moon, S.C. Chang, C.S. Huang, R.F. Chang, Breast tumor classification using fuzzy clustering for breast elastography, Ultrasound Med. Biol. 37 (2011) 700–708. [18] J.S. Suri, Advances in Diagnostic and Therapeutic Ultrasound Imaging, Artech House, Boston; London, 2008. [19] R. Deriche, Fast algorithms for low-level vision, IEEE Trans. Pattern Anal. Mach. Intelligence 12 (1990) 78–87. [20] R. Malladi, J.A. Sethian, B.C. Vemuri, Shape modeling with front propagation - a level set approach, IEEE Trans. Pattern Anal. Mach. Intelligence 17 (1995) 158– 175. [21] S. Osher, J.A. Sethian, Fronts propagating with curvature-dependent speed algorithms based on Hamilton-Jacobi formulations, J. Comput. Phys. 79 (1988) 12–49. [22] S.C. Chang, Y.C. Lai, Y.H. Chou, R.F. Chang, Breast elastography diagnosis based on dynamic sequence features, Med. Phys. 40 (2013) 022905. [23] W.C. Shen, R.F. Chang, W.K. Moon, Computer aided classification system for breast ultrasound based on breast imaging reporting and data system (BIRADS), Ultrasound Med. Biol. 33 (2007) 1688–1698. [24] W.C. Shen, R.F. Chang, W.K. Moon, Y.H. Chou, C.S. Huang, Breast ultrasound computer-aided diagnosis using BI-RADS features, Academic Radiol. 14 (2007) 928–939. [25] W.K. Moon, C.S. Huang, W.C. Shen, E. Takada, R.F. Chang, J. Joe, et al., Analysis of elastographic and B-mode features at sonoelastography for breast tumor classification, Ultrasound Med. Biol. 35 (2009) 1794–1802. [26] Y.C. Chang, M.C. Yang, C.S. Huang, S.C. Chang, G.Y. Huang, W.K. Moon, et al., Automatic selection of representative slice from cine-loops of real-time sonoelastography for classifying solid breast masses, Ultrasound Med. Biol. 37 (2011) 709–718. [27] A. Goddi, M. Bonardi, S. Alessi, Breast elastography: a literature review, J. Ultrasound 15 (2012) 192–198. [28] J.M. Chang, W.K. Moon, N. Cho, S.J. Kim, Breast mass evaluation: factors influencing the quality of US elastography, Radiology 259 (2011) 59–64.

W.K. Moon et al. / Ultrasonics 78 (2017) 125–133 [29] C.C. Chang, C.J. Lin, LIBSVM: a library for support vector machines, ACM Trans. Intelligent Syst. Technol. 2 (2011). [30] E. Alpaydin, Introduction to Machine Learning, MIT Press, Cambridge, Mass., 2004. [31] H. Johansson, S. Gandini, B. Bonanni, F. Mariette, A. Guerrieri-Gonzaga, D. Serrano, et al., Relationships between circulating hormone levels, mammographic percent density and breast cancer risk factors in postmenopausal women, Breast Cancer Res. Treat. 108 (2008) 57–67. [32] A. Evans, P. Whelehan, K. Thomson, D. McLean, K. Brauer, C. Purdie, et al., Quantitative shear wave ultrasound elastography: initial experience in solid breast masses, Breast Cancer Res. 12 (2010) R104.

133

[33] K.S. Bhatia, C.C. Cho, C.S. Tong, E.H. Yuen, A.T. Ahuja, Shear wave elasticity imaging of cervical lymph nodes, Ultrasound Med. Biol. 38 (Feb 2012) 195– 201. [34] J.H. Yoon, H.K. Jung, J.T. Lee, K.H. Ko, Shear-wave elastography in the diagnosis of solid breast masses: what leads to false-negative or false-positive results?, Eur Radiol. 23 (2013) 2432–2440. [35] R.G. Barr, Z. Zhang, Shear-wave elastography of the breast: value of a quality measure and comparison with strain elastography, Radiology 275 (2015) 45– 53. [36] S.H. Lee, J.M. Chang, N. Cho, H.R. Koo, A. Yi, S.J. Kim, et al., Practice guideline for the performance of breast ultrasound elastography, Ultrasonography 33 (2014) 3–10.