Ultrasound in Med. & Biol., Vol. 42, No. 5, pp. 1211–1220, 2016 Copyright Ó 2016 World Federation for Ultrasound in Medicine & Biology Printed in the USA. All rights reserved 0301-5629/$ - see front matter
http://dx.doi.org/10.1016/j.ultrasmedbio.2015.12.015
d
Original Contribution BREAST DENSITY ANALYSIS WITH AUTOMATED WHOLE-BREAST ULTRASOUND: COMPARISON WITH 3-D MAGNETIC RESONANCE IMAGING JEON-HOR CHEN,*y YAN-WEI LEE,z SI-WA CHAN,x{ DAH-CHERNG YEH,k and RUEY-FENG CHANGz{ * Tu & Yuen Center for Functional Onco-Imaging, Department of Radiological Sciences, University of California, Irvine, California, USA; y Department of Radiology, E-Da Hospital and I-Shou University, Kaohsiung, Taiwan; z Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan; x Department of Radiology, Taichung Veterans General Hospital, Taichung, Taiwan; { Graduate Institute of Biomedical Electronics and Bioinformatics, National Taiwan University, Taipei, Taiwan; and k Breast Center, Taichung Veterans General Hospital, Taichung, Taiwan (Received 6 August 2015; revised 28 October 2015; in final form 16 December 2015)
Abstract—In this study, a semi-automatic breast segmentation method was proposed on the basis of the rib shadow to extract breast regions from 3-D automated whole-breast ultrasound (ABUS) images. The density results were correlated with breast density values acquired with 3-D magnetic resonance imaging (MRI). MRI images of 46 breasts were collected from 23 women without a history of breast disease. Each subject also underwent ABUS. We used Otsu’s thresholding method on ABUS images to obtain local rib shadow information, which was combined with the global rib shadow information (extracted from all slice projections) and integrated with the anatomy’s breast tissue structure to determine the chest wall line. The fuzzy C-means classifier was used to extract the fibroglandular tissues from the acquired images. Whole-breast volume (WBV) and breast percentage density (BPD) were calculated in both modalities. Linear regression was used to compute the correlation of density results between the two modalities. The consistency of density measurement was also analyzed on the basis of intraand inter-operator variation. There was a high correlation of density results between MRI and ABUS (R2 5 0.798 for WBV, R2 5 0.825 for PBD). The mean WBV from ABUS images was slightly smaller than the mean WBV from MR images (MRI: 342.24 ± 128.08 cm3, ABUS: 325.47 ± 136.16 cm3, p , 0.05). In addition, the BPD calculated from MR images was smaller than the BPD from ABUS images (MRI: 24.71 ± 15.16%, ABUS: 28.90 ± 17.73%, p , 0.05). The intra-operator and inter-operator variant analysis results indicated that there was no statistically significant difference in breast density measurement variation between the two modalities. Our results revealed a high correlation in WBV and BPD between MRI and ABUS. Our study suggests that ABUS provides breast density information useful in the assessment of breast health. (E-mail: rfchang@ csie.ntu.edu.tw) Ó 2016 World Federation for Ultrasound in Medicine & Biology. Key Words: Automated whole-breast ultrasound, Rib shadow, Chest wall segmentation, Percentage breast density, Fuzzy C-means.
INTRODUCTION
tool used to assess breast density; however, tissue overlap is one of the major disadvantages of mammographic images and makes the accurate evaluation of whole-breast volume (WBV) and fibroglandular tissue volume (FV) difficult. In recent years, digital breast tomosynthesis (DBT), which is a pseudo-3-D examination of the breast, has gradually gained popularity in clinical breast imaging. In addition to its use in cancer diagnosis, DBT has also been investigated for its potential value in the measurement of breast density. The methodologic approaches to analysis of breast density, however, have not yet been clarified (Bakic et al. 2009; Tagliafico et al. 2012). Two automated breast assessment tools have been approved by the U.S. Food and Drug Administration and are increasingly being used. One is Quantra (Ciatto et al. 2012), and the other is Volpara (Seo et al.
Breast cancer, one of the most common malignancies, is a major cause of death among women worldwide. Many studies (Boyd et al. 2011; Freer 2015; Johansson et al. 2008; Li et al. 2012; McCormack and dos Santos Silva 2006; Shepherd et al. 2011; Varghese et al. 2012) have reported that women with dense breasts are at a higher risk for breast cancer than women whose breasts are less dense. Hence, breast density is known to be a clinically highly significant predictor of breast cancer risk. Conventionally, 2-D mammography is the main imaging
Address correspondence to: Ruey-Feng Chang, Department of Computer Science and Information Engineering, National Taiwan University, Taipei, 10617, Taiwan ROC. E-mail:
[email protected] 1211
1212
Ultrasound in Medicine and Biology
2013); both tools provide an objective estimate of the total volume of fibroglandular tissue, as well as the percentage breast density. Despite these recent developments, because the analysis is based on 2-D projection acquisition, these tools still suffer from the tissue overlap problem. The accuracy of breast density determined by mammography is of serious concern and likely to be inaccurate (Kopans 2008). Different from DBT, magnetic resonance imaging (MRI) and 3-D automated whole-breast ultrasound (ABUS) provide true volumetric imaging data for density analysis. Three-dimensional breast MRI provides better resolution of soft tissues, which helps to distinguish different organs/tissues in the images. ABUS renders breast echoic reflex images to construct breast tissue texture and, furthermore, establish whole-breast structure. Skin, subcutaneous fat, fibroglandular tissue, nipple, retromammary fat, muscle fascia and ribs can be observed in ABUS images (Moon et al. 2011). The echogenicity of fibroglandular tissue is higher than that of subcutaneous and retromammary fat. Three-dimensional MRI and 3-D ABUS have been used in recent studies (Moon et al. 2011; Nie et al. 2008) to assess breast density, and more reasonable consistency has been achieved. Hence, the two 3-D imaging modalities, breast MRI and ABUS, have the potential to more accurately calculate breast density. Breast density represents the total area of fibroglandular tissue over the whole breast region. The breast is located in a region containing different tissue types, including skin, subcutaneous fat, breast parenchyma, retromammary fat, pectoral muscles and ribs, which means that removal of non-breast regions is necessary. The conventional methods (Klifa et al. 2010; Li et al. 2012; Zhou et al. 2001) used to measure breast density are interactive. Interactive segmentation and thresholding have been used to separate breast regions and fibroglandular tissue regions. Many previous studies (Clendenen et al. 2013; Klifa et al. 2004; 2007; Li et al. 2012; Moon et al. 2011; van Engeland et al. 2006) used handheld segmentation to define the breast region; thus, the WBV depends on the operator’s control. In ABUS images, the ultrasound reflects back different pixel values depending on the tissue hardness and produces hyper-echoic and hypo-echoic regions. The rib shadow is produced by the reflection of ultrasound on ribs. Rib shadow can provide useful information for segmenting breast regions. In this study, a semi-automatic breast segmentation method based on 3-D ABUS images and using the rib shadow to extract breast region and calculate breast density is proposed. We used this semiautomatic segmentation method to maintain consistency in WBV. We used 3-D MRI density as the ground truth to validate the results. A flowchart of the overall analysis is provided in Figure 1.
Volume 42, Number 5, 2016
METHODS Patients We collected imaging data from 32 women. Each case had both 3-D ABUS images and 3-D MRI images. Nine women were excluded from the study because of an unclear rib shadow region on 3-D ABUS. In this study, we analyzed 46 normal breasts for breast density measurement in 23 women. The mean age of the 23 patients was 40.5 y, with a standard deviation (SD) of 8.2 y. None of the patients had a known history of breast disease. Informed consent was obtained from each patient to participate in this study, which was approved by the ethics committee of Taichung Veterans General Hospital. Imaging acquisition Three-dimensional ABUS images were acquired with an ACUSON S2000 Automated Breast Volume Scanner (Siemens Medical Solutions, Mountain View, CA, USA). Images were obtained from automated scanning using a wide-bandwidth (maximum field of view: 154 3 168 mm) linear transducer at 10-MHz gray-scale frequency. Images were constructed from 2-D slices from the transverse view (maximum resolution: 565 3 719 pixels with 318 frames). Each breast was covered by two passes of ABUS, and the nipple was included in both passes. We cut the overlap region using the vertical line across the nipple as reference. Three-dimensional MRI images were acquired with a Philips Achieva 3.0-T scanner (Eindhoven, Netherlands) using the turbo spin-echo T1-weighted pulse sequence without fat suppression. Images were obtained from 90 image slices 2 mm in thickness covering the whole breast (TR/TE 5 645/9.0 ms, echo
Fig. 1. Flowchart of overall analysis of ABUS and MRI results. ABUS 5 automated whole-breast ultrasound; MRI 5 magnetic resonance imaging.
US/MRI comparison in breast density analysis d J.-H. CHEN et al.
train 5 5, slice gap 5 0, phase encoding R-L, bandwidth per pixel 5 174 Hz, field of view 5 330 mm, imaging matrix 5 328 3 384 and parallel imaging with SENSE factor 5 2). MRI segmentation We employed the template-based automatic segmentation method proposed by Lin et al. (2013) to exclude the chest region from MRI data. This method manually delineates the contour of the chest and marks three body landmarks (thoracic spine and bilateral boundaries of the pectoral muscle) to produce a template image used to register other images. When iterated registration was applied to all images, the three-point V-shape cut and the lung contour helped to exclude the chest region. After the procedure, the whole breast region was obtained to calculate WBV, and the fuzzy C-means (FCM) cluster method was used to segment the fibroglandular tissue and acquire BPD. For details on the methodology, refer to Lin et al. (2013). In this study, we preset the WBV on MRI images, which were used as the ground truth for density measurements with ABUS. ABUS segmentation In this section, we propose a semi-automatic ABUS segmentation method based on rib shadow to remove non-breast regions. Figure 2 illustrates the visible breast and non-breast structures on the ABUS sagittal view, from the top of the image slice to the bottom, including skin, breast, pectoral muscles, ribs, rib shadow and intercostal space. We calculated the WBV with the remaining the breast regions and eliminated the other non-breast regions. Pre-processing Generally, ABUS images were scanned from regions below the collar bones to the nipples, and the images were shown in the transverse view. In Figure 3a is the transverse view slice, which has no clear texture
1213
and in which it was difficult to identify different tissues to segment the breast region. The sagittal view ultrasound images provide abundant information to recognize the different tissue types and to segment the chest wall. Hence, we transformed the transverse view image into a sagittal view first. Figure 3b is the sagittal view slice. From this view, ribs, rib shadow and different tissues can be recognized clearly. Rib shadow is the dark region below the ribs through which no ultrasound waves pass; hence, rib shadow had a lower pixel value. We enhanced the image contrast using the sigmoid filter (Braun and Fairchild 1999), so the difference between the rib shadow area with decreased pixel intensity value and other tissues with increased pixel intensity values was more marked. The sigmoid filter computed the sigmoid function pixelwise and was defined as 1 1Imin SðxÞ 5 ð Imax 2Imin Þ 3 x2b 11e2 a
(1)
where x is the input pixel, and S(x) is the output pixel. Imax and Imin are the maximum pixel and minimum pixel in the image, respectively. a and b are constant parameters to control the enhancement level. After contrast enhancement, we used the anisotropic diffusion filter (Weickert 1998) to enhance the image. Mathematically, the diffusion process is defined as v Iðx; y; tÞ 5 div½ðckVIkÞ$VI vt
(2)
where div is the divergence operator, V denotes the gradient operator, jj$jj denotes the magnitude, I(x,y,0) is the initial image, x and y are the image coordinates, t is the iterator step and c (jjVIjj) is the diffusion function, which is a non-negative monotonically decreasing function depending on the image gradient magnitude. To encourage smoothing within a region and discourage smoothing across region boundaries, the diffusion function was calculated as
Fig. 2. Sagittal view 3-D automated whole-breast ultrasound image describes the visible structure of the breast (from top to bottom including skin, breast, pectoral muscles, ribs, rib shadow and intercostal space).
1214
Ultrasound in Medicine and Biology
Volume 42, Number 5, 2016
Fig. 3. (a) Transverse view automated whole-breast ultrasound slice. (b) Sagittal view automated whole-breast ultrasound slice. (c) Image after pre-processing (sigmoid image filter and anisotropic diffusion filter). The darker area (rib shadow) preserves the local information, and the edge information was protected (sigmoid filter: a 5 50, b 5 170, anisotropic diffusion filter parameters: time step 5 0.1, iterations 5 30, conductance 5 10).
" 2 # kVIk ðckVIkÞ 5 EXP 2 K
(3)
where K is the diffusion coefficient. In this study, the value of K was set at 1. On one hand, ultrasound images have speckle noise caused by the destructive and constructive interference by the radiation from the scattering elements in the tissue. The anisotropic diffusion filter can remove the speckle noise. On the other hand, rib shadow areas appear gray, blocklike and uneven. The anisotropic diffusion filter not only made the rib shadow to appear as dark block areas, but also avoids the major drawback of conventional spatial filters and improves image quality while preserving important edge
information. In our experiment, the rib shadow contrast was enhanced and speckle noise was reduced after applying the sigmoid filter and anisotropic diffusion filter, as illustrated in Figure 3c. Extraction of chest wall reference line In this method, we combined the local information and global information to segment the chest wall line. The local information was extracted from a single slice, and the global information came from the whole set of images. After pre-processing, we obtained a highcontrast, de-noised and blurry image. The edge information was preserved and that helped us to execute the chest wall line extraction process. We applied Otsu’s
US/MRI comparison in breast density analysis d J.-H. CHEN et al.
thresholding method to the image after pre-processing to separate the rib shadow and perform clusteringbased image thresholding automatically, and separated the image into foreground (rib shadow) and background (other tissue) components. The threshold value was the maximum between-class variance. Figure 4 is the image after application of Otsu’s thresholding method, revealing the rib shadow separated from the image, which is called the rib map. Each slice extracted the local rib shadow information, which included the local rib location, its height and its width. The areas between the ribs (intercostal space), which do not belong to the breast region and should be removed, can be seen in Figure 4. From the rib shadow information, we can observe the trend of the whole chest cage. However, each slice provides only a fragment of rib shadow information; thus, the global information is needed. The mean projection method projects the pixels on the same location crossing through different slices along an axis and counts the mean intensity value of the pixels at the same location through all slices. Pixels were preserved if the pixel intensity value was greater than the mean count; otherwise, the pixel was discarded and its value was set to zero in the same location across all slices. All rib shadow information via the mean projection method was fused on a plane, and all of the fragment rib shadow information was combined to generate the global chest information, which describes the profile of the chest cage and rib distribution. Figure 5 illustrates the result after use of the mean projection filter. We extracted the dividing line from the reference line of the chest wall. In Figure 5, the rib shapes appear cagelike, and when all the slices were projected on an axis plane, the distribution of the rib location area was shown. The white region represents all of the rib locations in the chest, describing the rib location intersection region (all of the ribs are included in this region, the global rib shadow location information was included). The remaining black region is the breast region, which describes the distribution region of the whole breast tissue in the images.
Fig. 4. Rib map, results processed after Otsu’s thresholding method. Rib shadows and other darker regions were separated into foreground (white) and background (black) areas.
1215
Fig. 5. Chest wall reference line. The foreground (white) region describes the rib information, including rib location, rib height and rib distribution. The background (black) region describes the average proportion of whole breast tissue on images.
Removal of chest region and skin After the aforementioned procedures, we obtained an image with the global information and all of the local information slices. We combined all information to segment the chest wall. In Figure 4, we can observe the intercostal spaces between ribs, which do not belong to the breast region. Afterward, we used the chest wall reference line to eliminate intercostal spaces. The chest wall reference line will be referenced by each slice to find the divided line which helps us to remove the chest region. At first, we recorded the highest location of the rib shadow in the rib map slice. The highest location of the rib shadow was representative, clear and evident in the image slice that describes the smallest breast thickness in the slice. Second, we extracted the coordination from the highest point in the highest rib shadow, recorded the row number of the coordination from the rib map and found the pixel location in the corresponding row number in the chest wall line. Third, we aligned the chest wall reference line with the corresponding row number and transplanted the chest wall line pixel on each rib map slice. This step combined the global information with the local information for each slice. Fourth, so as not to affect the shape of the original rib map, we regulated the height of the chest wall reference line and reserved the length of the half-rib diameter (Mohr et al. 2007) (about 5.5–7.0 mm) to maintain the shape of the chest contour, so we regarded the region below this line as chest region and cut the pectoral muscles last. The average thickness of pectoral muscles (Koo et al. 2010) in an adult is 1.5–2.0 cm; we cut the thickness of pectoral muscles based on the assumption of Koo and co-workers. This line combined the global information and local information to segment the breast; the region above the line was the breast region, and that below the line was the chest region. We called this line the chest wall line. The segmentation results are illustrated in Figure 6, after combination of the rib shadow information and the cutting of pectoral muscles. The outmost 12–18 pixels (in total 1.5–1.7 mm) of the ABUS images was regarded as skin region and cut off (Moon et al. 2011). After the segmentation process, we retained the breast region to calculate the WBV and BPD. Because each breast was covered by two passes
1216
Ultrasound in Medicine and Biology
of images, we used a vertical line across the nipple in longitudinal view to manually cut the overlapping region. Figure 7 illustrates the manual cutting process to remove the overlapping region using the nipple location as a reference. Breast percentage density calculation After breast segmentation, we calculated the WBV with different modalities (ABUS and MRI). We employed the FCM classifier to separate the fibroglandular tissue and fatty tissue and measured the BPD (Nie et al. 2008). The FCM classifier provided a pixel-based segmentation with clusters that are defined by their centroids (Bezdek et al. 1984). The values of cluster centroids are updated iteratively, and the pixels are classified into different clusters according to the probability of each centroid value. The equations for the FCM classifier are PN
m i 5 1 uij $xi
c j 5 PN
m i 5 1 uij
; uij 5
Pc k 5 1
Volume 42, Number 5, 2016
was determined, the FV and BPD could be calculated. BPD was defined as BPDmodalities 5
fibroglandular volumemodalities $100% (5) total breast volumemodalities
where fibroglandular volumemodalities is FV, total breast volumemodalities is the WBV and BPDmodalities is the percentage of FV in the WBV with different modalities. The segmentation and FCM results are illustrated in Figure 8. Statistical analysis We compared the ABUS and MRI results using linear regression analysis, calculated the correlation factor and fitted a line to observe their relationship. Linear regression analysis can indicate the relevance, related direction and strength, and produce a quantifiable regres-
XN XC 2 1 m u x i 2cj 2 ; J 5 x 2c m21 ij i 5 1 j 5 1 k i jk
(4)
kxi 2ck k
which compute the object function J, where N is the number of pixels, m is a user-defined variable, uij is the coefficient value of pixel xi to centroid cj and C is the number of clusters. For our experiments, the operator chose three or four clusters to separate the breast into several part regions, and the fibroglandular tissue was assigned as the bright region (dense region). After the dense region
Fig. 6. (a) Alignment of chest wall reference line. (b) Cutting of pectoral muscles. (c) Region removed below chest wall line.
sion function (Draper et al. 1966). To test measurement consistency, we analyzed the inter-operator and intraoperator variation of the WBV and BPD values acquired for nine randomly selected breasts imaged with the two different modalities. An operator analyzed these nine
Fig. 7. Manual cutting of the overlap region using a vertical line across the nipple as reference.
US/MRI comparison in breast density analysis d J.-H. CHEN et al.
1217
trials. The last two were chosen by operators 2 and 3, respectively.
Fig. 8. Breast segmentation results. (a) The segmentation results with the region containing the breast and without skin. (b) Fuzzy C-means classifier results (cluster number 5 3) revealing separation of fibroglandular tissue and fatty tissue.
breasts three times, and the other two operators analyzed the same nine breasts once. The coefficient of variation (CV, defined as the ratio of the SD to the mean) was used as an indicator of measurement consistency. The p-value was also calculated to assess the level of significance. A p-value , 0.05 was defined as indicating statistical significance. Bland–Altman difference plots were used to assess differences in measurements between the two imaging modalities. The three horizontal lines in the plot represent the mean value and mean value 6 2 * SD. RESULTS Intra-operator and inter-operator variation analysis The calculated mean and standard variation values for WBV and BPD are listed in Table 1. In intraoperator analysis, the CV of WBV was 2.24% (1.29%– 3.09%) and the CV of BPD was 3.28% (0.98%–6.11%). In inter-operator analysis, the CV of WBV was 3.13% (1.82%–5.71%) and CV of BPD was 4.41% (1.70%– 6.62%). In cases 3 and 7 in Table 1, the operators chose the different FCM cluster numbers, resulting in higherthan-average CVs (8.28% and 4.98%, respectively). The FCM cluster number was selected by the operator. The first three were chosen by operator one in three
Comparison of density results between modalities The linear regression of the density result is illustrated in Figure 8. The WBV values (mean 6 SD) for MRI and ABUS images were 344.06 6 127.89 and 325.46 6 136.16 cm3, respectively. Overall, the WBV values from ABUS images were smaller than those from MRI images. In breast volume analysis, the correlation factor (r) was 0.89 (p , 0.05). The FV values for MRI and ABUS images were 86.53 6 68.68 and 94.32 6 72.51 cm3, respectively. The results indicate that the FV values from ABUS images were larger than those from MRI images. In FV analysis, r was 0.90 (p . 0.05). The BPD values calculated from ABUS images were larger than those calculated from MRI images. The BPD values were 24.81 6 15.64% for MRI images and 29.32 6 17.86% for ABUS images. In BPD analysis, r was 0.91. The difference between these two modalities was not statistically significant (p . 0.05). The linear regression between these two modalities is illustrated in Figure 9(a). Bland–Altman difference plots for WBV, FVand BPD values from MRI and ABUS images are provided in Figure 9(b). Note that the operators who performed and analyzed the ABUS data were blind to the MRI results. DISCUSSION AND CONCLUSIONS Automated whole-breast ultrasound has been reported to be a useful modality for detecting earlystage breast cancer (Corsetti et al. 2008). In recent years, ABUS has been widely used to screen women at risk of breast cancer (Lord et al. 2007). Its non-invasiveness and real-time display are its advantages. As an adjunct to mammography, ABUS has been used in clinical diagnosis, especially in women with dense breast tissue
Table 1. Inter-operator and intra-operator variation analysis ABUS breast volume (cm3) Case no. 1 2 3 4 5 6 7 8 9 CV
Intra-operator
Inter-operator
364.98 6 8.63* (2.36%) 479.45 6 6.19 (1.29%) 398.65 6 10.94 (2.74%) 584.27 6 8.43 (1.44%) 416.31 6 9.22 (2.21%) 318.66 6 8.35 (2.62%) 401.12 6 7.22 (1.80%) 296.34 6 9.15 (3.09%) 371.83 6 9.64 (2.59%) 2.24%
368.56 6 8.24 (2.24%) 470.69 6 11.37 (2.42%) 421.67 6 12.83 (3.04%) 569.37 6 10.36 (1.82%) 429.44 6 16.68 (3.88%) 322.97 6 18.46 (5.72%) 413.64 6 10.67 (2.58%) 308.94 6 10.35 (3.35%) 367.15 6 11.54 (3.14%) 3.13%
ABUS breast density (%) FCM cluster no. 3,3,3,3,3 4,4,4,4,4 3,3,3,4,3 3,3,3,3,3 3,3,3,3,3 3,3,3,3,3 4,4,4,4,3 4,4,4,4,4 3,3,3,3,3
Intra-operator
Inter-operator
10.47 6 0.64 (6.11%) 17.64 6 0.31 (1.76%) 14.19 6 0.57 (4.02%) 36.82 6 0.36 (0.98%) 27.66 6 0.87 (3.15%) 22.85 6 0.74 (3.24%) 29.81 6 0.95 (3.19 %) 13.51 6 0.61 (4.52%) 19.90 6 0.52 (3.61%) 3.28%
10.12 6 0.73 (7.21%) 16.48 6 0.34 (2.06%) 11.60 6 0.96 (8.28%) 39.45 6 0.67 (1.70%) 26.47 6 0.94 (3.55%) 22.34 6 0.88 (3.94%) 30.73 6 1.53 (4.98%) 14.08 6 0.68 (4.83%) 20.14 6 0.63 (3.13%) 4.41%
ABUS 5 automated whole-breast ultrasound; CV 5 coefficient of variation; FCM 5 fuzzy C-means classifier. * Mean 6 SD.
1218
Ultrasound in Medicine and Biology
Volume 42, Number 5, 2016
Fig. 9. (a) Linear regression of breast volume, fibroglandular volume and percentage density between MRI ABUS. (b) Bland–Altman difference plot of differences in breast volume, fibroglandular volume and percentage density between MRI and ABUS images. ABUS 5 automated whole-breast ultrasound; MRI 5 magnetic resonance imaging.
(Kelly et al. 2010). It has been proven by many studies that, for diagnostic purposes, MRI is more sensitive than sonography for the detection of small breast lesions (Kuhl et al. 2005; Warner et al. 2001). With the introduction of ABUS, diagnostic sensitivity has been improved, but is still inferior to that of MRI (Chae et al. 2013). In this study, the goal was to compare breast density measured with a new sonographic method with that measured with MRI, which is used as the ground truth. With both modalities, only normal breast tissue is targeted. The diseased breast with cancer is normally not measured because the space-occupying lesion itself is not normal tissue and has displaced or destroyed the normal breast tissue, making the density measurement inaccurate. In the studies of Moon et al. (2011; 2013), ABUS images were employed to assess BPD, which is regarded as a risk factor to assess the healthy level of the breasts. However, to date there is no efficient segmentation method for separating the breast region on ABUS images. In this study, we proposed a semi-automatic chest wall segmentation method based on rib shadow, which separates the breast region in ABUS images. After segmentation, we used the FCM classifier to separate fibroglandular tissue from breast fatty regions, then calculated the 3-D breast volume and compared the acquired BPD with that from MRI. Klifa et al. (2004) compared density measurement variation among manual,
global threshold and FCM and noted that the FCM classifier had less user variation than other methods. However, choosing different cluster numbers might have affected the segmented FV. Our experiment results indicated that different cluster numbers led to different density results. Breast percentage density is calculated as the volume proportion of fibroglandular tissue divided by WBV. Accurate and consistent delineation of the breast boundary is crucial. In this study, we used the physical features of rib shadow imaging on ABUS to define the breast boundary. In some proposed methods (Chang et al. 2008; Kim et al. 2015; Tan et al. 2013; 2014), ribs and rib shadow were excellent features for depicting the chest region or chest wall line. The advantage of breast segmentation using rib shadow is that the ribs are easy to identify in ABUS images and rib shadow indicates the location of ribs, which helps us to determine the location of the line separating the chest region and breast region. The interference by rib shadow was the disadvantage of the rib shadow segmentation method. If rib shadow is not intact or the tumor location is connected to the chest wall line, oversegmentation will occur. Thus, clear definition of the rib shadow is necessary in our method. Our segmentation method based on rib shadow, combining local information produced by Otsu’s method with global information produced by mean projection performed well for 23
US/MRI comparison in breast density analysis d J.-H. CHEN et al.
women and revealed a high correlation of BV with MRI results. Nevertheless, nine women were excluded from the study because of indistinct rib shadow regions. The results from our study indicate that BVs measured from MRI images are larger than BVs measured from ABUS images. Among the reasons for the different measurement results, two are discussed here: The first is the difference in image acquisition method. MRI scans produce serial images, but ABUS scans are covered by two passes of the images. It was not known with certainty whether the whole breast was scanned in the two passes. Moreover, some regions in ABUS images might be overcut or undercut in our cutting process to combine two passes, which would lead to calculation errors. In future work, we will combine and register the rib shadow information from the two passes of the images to prevent the operator from miscutting. The second reason is the scanning position in different modalities. The patients were supine during ABUS scanning and were prone in MRI scanning. Thus, in MRI studies, because of gravity, more breast volume might be scanned and measured. The results of our study indicate that FV and BPD measured with ABUS were higher than FV and BPD acquired with MRI. The reason may be the different gray values in the dense tissue between the two modalities, resulting in misclassification errors. MRI and ABUS use different imaging mechanisms and different scan positions to generate images, which might affect glandular pixel values or might produce different fibroglandular tissue pixel values, or some pixel value will be missed in the same area. ABUS measured lower BVs and higher FVs, which produced higher BPDs. The reproducibility test for ABUS density measurements revealed that the CVs for both intra-operator and inter-operator variation analysis were low (,5%), indicating that ABUS can be a new modality for breast density analysis. Both cases 3 and 7 (in Table 1) had indistinct fibroglandular tissue in the breast region that led the operators to choose different cluster numbers and resulted in CVs larger than those for the other women. The usefulness of rib shadow may be limited if ABUS images have insufficient scan depth to depict the ribs and to produce the chest wall reference line. Moreover, with ABUS, the level of compression on the breast may affect the volume measurement. The relation between compression level and breast volume is worthy of further investigation. CONCLUSIONS We have proposed a segmentation method based on rib shadow that generates a chest wall reference line, which is used to separate the breast region from breast
1219
MR images. The ABUS results were well correlated with MRI breast density. Our study indicates that ABUS can provide breast density information useful in assessing breast health. Acknowledgments—The authors thank the Ministry of Science and Technology (MOST 103-2221-E-002-170-MY3) and Ministry of Education (AE-00-00-06) of the Republic of China for financial support. The authors also thank Rita Chang for spending time to edit the article.
REFERENCES Bakic PR, Carton AK, Kontos D, Zhang C, Troxel AB, Maidment AD. Breast percent density: Estimation on digital mammograms and central tomosynthesis projections. Radiology 2009;252:40–49. Bezdek JC, Ehrlich R, Full W. Fcm: The fuzzy c-means clustering algorithm. Computers Geosci 1984;10:191–203. Boyd NF, Martin LJ, Yaffe MJ, Minkin S. Mammographic density and breast cancer risk: Current understanding and future prospects. Breast Cancer Res 2011;13:223. Braun GJ, Fairchild MD. Image lightness rescaling using sigmoidal contrast enhancement functions. J Electronic Imaging 1999;8: 380–393. Chae EY, Shin HJ, Kim HJ, Yoo H, Baek S, Cha JH, Kim HH. Diagnostic performance of automated breast ultrasound as a replacement for a hand-held second-look ultrasound for breast lesions detected initially on magnetic resonance imaging. Ultrasound Med Biol 2013;39:2246–2254. Chang RF, Shen YW, Chen J, Chou YH, Huang CS. Rib detection for whole breast ultrasound image. In: Giger ML, Karssemeijer N, (eds). Medical Imaging 2008: Computer-Aided Diagnosis, 6915. San Diego, CA: Proc SPIE; 2008. p. 691525. Ciatto S, Bernardi D, Calabrese M, Durando M, Gentilini MA, Mariscotti G, Monetti F, Moriconi E, Pesce B, Roselli A. A first evaluation of breast radiological density assessment by quantra software as compared to visual classification. Breast 2012;21:503–506. Clendenen TV, Zeleniuch-Jacquotte A, Moy L, Pike MC, Rusinek H, Kim S. Comparison of 3-point dixon imaging and fuzzy C-means clustering methods for breast density measurement. J Magn Reson Imaging 2013;38:474–481. Corsetti V, Houssami N, Ferrari A, Ghirardi M, Bellarosa S, Angelini O, Bani C, Sardo P, Remida G, Galligioni E. Breast screening with ultrasound in women with mammography-negative dense breasts: Evidence on incremental cancer detection and false positives, and associated cost. Eur J Cancer 2008;44:539–544. Draper NR, Smith H, Pownell E. Applied regression analysis. New York: Wiley; 1966. Freer PE. Mammographic breast density: Impact on breast cancer risk and implications for screening. RadioGraphics 2015;35:302–315. Johansson H, Gandini S, Bonanni B, Mariette F, Guerrieri-Gonzaga A, Serrano D, Cassano E, Ramazzotto F, Baglietto L, Sandri MT. Relationships between circulating hormone levels, mammographic percent density and breast cancer risk factors in postmenopausal women. Breast Cancer Res Treat 2008;108:57–67. Kelly KM, Dean J, Comulada WS, Lee SJ. Breast cancer detection using automated whole breast ultrasound and mammography in radiographically dense breasts. Eur Radiol 2010;20:734–742. Kim H, Kim H, Hong H. Chest wall segmentation in automated 3-D breast ultrasound using rib shadow enhancement and multi-plane cumulative probability enhanced map. In: Hadjiiski LM, Tourassi GD, (eds). Medical Imaging 2015: Computer-Aided Diagnosis, 9414. Orlando, FL: Proc SPIE; 2015. p. 941423. Klifa C, Carballido-Gamio J, Wilmes L, Laprie A, Lobo C, DeMicco E, Watkins M, Shepherd J, Gibbs J, Hylton N. Quantification of breast tissue index from mr data using fuzzy clustering. In: Proceedings, IEMBS’04: 26th Annual IEEE International Conference of the Engineering in Medicine and Biology Society. New York: IEEE, 2004:1667–1670. Klifa C, Carballido-Gamio J, Wilmes L, Laprie A, Shepherd J, Gibbs J, Fan B, Noworolski S, Hylton N. Magnetic resonance imaging for
1220
Ultrasound in Medicine and Biology
secondary assessment of breast density in a high-risk cohort. Magn Reson Imaging 2010;28:8–15. Klifa C, Hattangadi J, Watkins M, Li A, Sakata T, Tromberg B, Hylton N, Park C. Combination of magnetic resonance imaging and diffuse optical spectroscopy to predict radiation response in the breast: An exploratory pilot study. In: Azar FS, (ed). Multimodal Biomedical Imaging II: Biomedical Optics (BiOS) 2007, 64310. San Jose, CA: Proc SPIE; 2007. p. 64310C. Koo TK, Wong C, Zheng Y. Reliability of sonomyography for pectoralis major thickness measurement. J Manip Physiol Therap 2010;33: 386–394. Kopans DB. Basic physics and doubts about relationship between mammographically determined tissue density and breast cancer risk. Radiology 2008;246:348–353. Kuhl CK, Schrading S, Leutner CC, Morakkabati-Spitz N, Wardelmann E, Fimmers R, Kuhn W, Schild HH. Mammography, breast ultrasound, and magnetic resonance imaging for surveillance of women at high familial risk for breast cancer. J Clin Oncol 2005;23:8469–8476. Li J, Szekely L, Eriksson L, Heddson B, Sundbom A, Czene K, Hall P, Humphreys K. High-throughput mammographic-density measurement: A tool for risk prediction of breast cancer. Breast Cancer Res 2012;14:R114. Lin M, Chen JH, Wang X, Chan S, Chen S, Su MY. Template-based automatic breast segmentation on MRI by excluding the chest region. Med Phys 2013;40:122301. Lord S, Lei W, Craft P, Cawson J, Morris I, Walleser S, Griffiths A, Parker S, Houssami N. A systematic review of the effectiveness of magnetic resonance imaging (MRI) as an addition to mammography and ultrasound in screening young women at high risk of breast cancer. Eur J Cancer 2007;43:1905–1917. McCormack VA, dos Santos Silva I. Breast density and parenchymal patterns as markers of breast cancer risk: A meta-analysis. Cancer Epidemiol Biomarkers Prev 2006;15:1159–1169. Mohr M, Abrams E, Engel C, Long WB, Bottlang M. Geometry of human ribs pertinent to orthopedic chest-wall reconstruction. J Biomech 2007;40:1310–1317. Moon WK, Lo CM, Chang JM, Bae MS, Kim WH, Huang CS, Chen JH, Kuo MH, Chang RF. Rapid breast density analysis of partial volumes of automated breast ultrasound images. Ultrason Imaging 2013;35:333–343. Moon WK, Shen YW, Huang CS, Luo SC, Kuzucan A, Chen JH, Chang RF. Comparative study of density analysis using auto-
Volume 42, Number 5, 2016 mated whole breast ultrasound and MRI. Med Phys 2011;38: 382–389. Nie K, Chen JH, Chan S, Chau MKI, Hon JY, Bahri S, Tseng T, Nalcioglu O, Su MY. Development of a quantitative method for analysis of breast density based on three-dimensional breast MRI. Med Phys 2008;35:5253–5262. Seo J, Ko E, Han B-K, Ko EY, Shin JH, Hahn SY. Automated volumetric breast density estimation: A comparison with visual assessment. Clin Radiol 2013;68:690–695. Shepherd JA, Kerlikowske K, Ma L, Duewer F, Fan B, Wang J, Malkov S, Vittinghoff E, Cummings SR. Volume of mammographic density and risk of breast cancer. Cancer Epidemiol Biomarkers Prev 2011;20:1473–1482. Tagliafico A, Tagliafico G, Astengo D, Cavagnetto F, Rosasco R, Rescinito G, Monetti F, Calabrese M. Mammographic density estimation: One-to-one comparison of digital mammography and digital breast tomosynthesis using fully automated software. Eur Radiol 2012;22:1265–1270. Tan T, Platel B, Mann RM, Huisman H, Karssemeijer N. Chest wall segmentation in automated 3-D breast ultrasound scans. Med Image Anal 2013;17:1273–1281. Tan T, van Zelst J, Zhang W, Mann RM, Platel B, Karssemeijer N. Chest-wall segmentation in automated 3-D breast ultrasound images using thoracic volume classification. In: Aylward S, Hadjiiski LM, (eds). Medical Imaging 2014: Computer-Aided Diagnosis, 9035. San Diego, CA: Proc SPIE; 2014. p. 90351Y. Van Engeland S, Snoeren PR, Huisman H, Boetes C, Karssemeijer N. Volumetric breast density estimation from full-field digital mammograms. IEEE Trans Med Imaging 2006;25:273–282. Varghese JS, Thompson DJ, Michailidou K, Lindstr€om S, Turnbull C, Brown J, Leyland J, Warren RM, Luben RN, Loos RJ. Mammographic breast density and breast cancer: Evidence of a shared genetic basis. Cancer Res 2012;72:1478–1484. Warner E, Plewes D, Shumak R, Catzavelos G, Di Prospero L, Yaffe M, Goel V, Ramsay E, Chart P, Cole D. Comparison of breast magnetic resonance imaging, mammography, and ultrasound for surveillance of women at high risk for hereditary breast cancer. J Clin Oncol 2001;19:3524–3531. Weickert J. Anisotropic diffusion in image processing. Stuttgart: Teubner; 1998. Zhou C, Chan HP, Petrick N, Helvie MA, Goodsitt MM, Sahiner B, Hadjiiski LM. Computerized image analysis: Estimation of breast density on mammograms. Med Phys 2001;28:1056–1069.