- Email: [email protected]
Vascular Morphology and Tortuosity Analysis of Breast Tumor Inside and Outside Contour by 3-D Power Doppler Ultrasound
Ultrasound in Med. & Biol., Vol. 38, No. 11, pp. 1859–1869, 2012 Copyright Ó 2012 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.2012.06.010
d
Original Contribution VASCULAR MORPHOLOGY AND TORTUOSITY ANALYSIS OF BREAST TUMOR INSIDE AND OUTSIDE CONTOUR BY 3-D POWER DOPPLER ULTRASOUND YEUN-CHUNG CHANG,* YAN-HAO HUANG,y CHIUN-SHENG HUANG,z and RUEY-FENG CHANGyx * Department of Medical Imaging, National Taiwan Univesity Hospital and National Taiwan University College of Medicine, Taipei, Taiwan; y Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan; z Department of Surgery, National Taiwan University Hospital and National Taiwan University College of Medicine, Taipei, Taiwan; and x Graduate Institute of Biomedical Electronics and Bioinformatics, National Taiwan University, Taipei, Taiwan (Received 20 December 2011; revised 16 June 2012; in final form 26 June 2012)
Abstract—This study aimed to evaluate morphologic and tortuous features of vessels inside and outside the tumor region on three-dimensional power Doppler ultrasonography (PDUS) in 113 breast mass lesions, including 60 benign and 53 malignant tumors. Compared with benign lesions, malignant breast lesions had significantly larger values of vascular morphologic and tortuous features and larger tumor sizes. The receiver operating characteristic curve analysis and Student’s t-test were used to estimate the performance of a proposed classification system using 13 vascular features and tumor size selected by the neural network. Accuracy, sensitivity, specificity, positive predictive value, negative predictive value and the AZ value of the diagnosis performance based on 14 features were 89.38% (101/113), 84.91% (45/53), 93.33% (56/60), 91.84% (45/49), 87.50% (56/64) and 0.9188, respectively. The three-dimensional PDUS morphologic and tortuous characteristics of blood vessels inside and outside breast mass lesions can be effectively used to classify benign and malignant tumors. (E-mail: [email protected]. edu.tw) Ó 2012 World Federation for Ultrasound in Medicine & Biology. Key Words: 3-D ultrasound, Breast tumor, Vascularity, Morphology, Tortuosity.
to differentiate between benign and malignant breast tumors (Hsiao et al. 2008; Sehgal et al. 2000). Vascular quantification using VI, flow index (FI) and vascularization flow index (VFI) from computer extracted volume-of-interest (VOI) and area outside the target lesion on 3-D PDUS can also differentiate between malignant and benign lesions (Huang et al. 2009). Moreover, vessel tortuosity, which is the estimation of the degree of vessel bending is considered to be related to intra-cranial malignancy and useful for assessing tumor activity (Bullitt et al. 2003, 2004, 2007). Compared with normal tissues, the vasculature of malignant lesion is more branching and tortuous and the vessels of malignant tumor are more chaotic and malformed than those of benign tumors (Huang et al. 2008). If the vessels are bent and twisted, the findings are usually associated with tumor malignancy (Chang et al. 2006, Chang et al. 2007). Combined morphologic and tortuous vascular features from 3-D PDUS have higher accuracy in differentiating benign from malignant masses than using 3-D VI alone (Chang et al. 2006, 2007). Good correlation of anatomic vascular distribution between 3-D PDUS and histopathology has also been reported (Strano et al. 2004). Doppler features suggestive
INTRODUCTION Angiogenesis in breast cancer is commonly assessed and quantified using microvessel density (MVD) (Uzzan et al. 2004). Color Doppler ultrasound vascularization index (VI), defined as the number of vessel pixels divided by the number of all pixels, reportedly correlates with MVD (Chen et al. 1999, 2002). Since vascular quantification from two-dimensional (2-D) ultrasound (US) may vary due to difference of selected planes, VI obtained from three-dimensional (3-D) power Doppler ultrasound (PDUS) is more reliable and reproducible for its capacity for volumetric coverage and global assessment of tumor vascularity (Ferrara et al. 1996; Krestan et al. 2002). Compared with 2-D US, 3-D color flow display provides stronger subjective appreciation of vascular morphology (Carson et al. 1997) and significantly higher correlation with MVD (Yang et al. 2002). Recent studies using 3-D PDUS show that vascularity surrounding the tumor and intra-tumor vascularity on power Doppler US are useful Address correspondence to: Ruey-Feng Chang, PhD, Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan 106, R.O.C. E-mail: [email protected] 1859
1860
Ultrasound in Medicine and Biology
Volume 38, Number 11, 2012
of malignant lesions include the presence of both peripheral and central vascularity, as well as penetrating and branching vessels (Lee et al. 2002; Strano et al. 2004). Analysis of vascular morphology and distribution is important to obtain spatial information of tumor heterogeneity and to differentiate breast tumors (Gokalp et al. 2009; LeCarpentier et al. 2008). The purpose of the current study was to assess the diagnostic performance of combined 3-D vascular morphologic and tortuous features in comparison to previous methods for discriminating benign from malignant tumors. The grey-scale tumor margin was used to categorize vascular distribution inside and outside the targeted mass, with quantification of vascular morphologic and tortuous features. MATERIALS AND METHODS Patients One hundred and thirteen (113) consecutive solid tumors (range, 0.2–4.0 cm; mean, 1.8 6 0.9 cm) that received 3-D PDUS before undergoing surgical excision or percutaneous needle biopsy due to suspicious US findings between June 2008 and June 2009 were evaluated after excluding nine patients (7%) with a tumor size larger than 4 cm, which could not be included in one 3-D acquisition. All benign lesions, histopathologically confirmed, were followed-up for stability for at least 2 years. Sixty were benign lesions, including nine fibroadenomas, 39 focal fibrocystic changes, five papillomas and seven nonspecified fibro-epithelial lesions. Fiftythree were malignant, including 34 invasive ductal carcinomas (IDC), 11 ductal carcinoma in situ (DCIS) and eight invasive lobular carcinomas. Informed consent was obtained from each patient prior to performing a biopsy. Institutional review board (IRB) approval was obtained for this study and informed consent was waived for this retrospective review. Ultrasound imaging All US images were acquired using a 3-D power Doppler US scanner (Voluson 730; GE Healthcare, Zipf, Austria) and 6–12 MHz dedicated volume transducer by three radiologists with 3–8 years of experience in breast ultrasound. A suitable volume box size of 4 3 4 cm2 with adjustable depth was used to include the lesion with appropriate surrounding tissue and vasculature and the sweep angle was 25 –30 . The 3-D volume files were then saved in Cartesian coordinates using the 4-DView program with the US scanner. Lesions larger than 4.0 cm were excluded. Imaging preprocessing and tumor segmentation In 3-D power Doppler data, the vessel image, which contains the information about the density of red blood
Fig. 1. Flow chart of the proposed computer-aided diagnosis (CAD) system.
cells and grey B-mode image were obtained simultaneously and saved in the same image file. Then, the vessel image and grey B-mode image were further divided into two separate image files to extract vessels and tumor information, respectively. The tumor region was segmented from B-mode grey-scale ultrasound images. Three 3-D image processing methods were adopted to reduce noises and enhance the contrast for tumor segmentation. A 3-D sigmoid filter (Suri 2008) was used to enhance the gaps of intensity distributions between objects, while the sigma edge-preserving filter (Lee 1983) was applied to smooth images and preserve the boundary information (Fig. 1). Finally, a gradient magnitude filter (Gonzalez and Woods 2002) was used to extract the input data of the segmentation method (Fig. 2). The tumor extraction algorithm was based on the 3-D level set method by which a stable contour was obtained after manually selecting a seed within tumor and iteratively evolving the isosurface (Malladi et al. 1995; Osher and Sethian 1988). The user reviewed all the 2-D slices extracted from the 3-D B-mode image. The tumor center on the slice containing the maximum tumor contour or dimension was selected as the seed. Additional seeds on other slices could be manually added for more precise contour if the tumor was not well segmented. In this study, the average number of seeds was 1.3 (range, 1–5 points). The segmented tumor region
Vascular morphology and tortuosity analysis d Y.-C. CHANG et al.
1861
Fig. 2. The process of segmentation of grey-scale ultrasound image. (a) Original grey-scale image. (b) Image on (a) after applying sigmoid filter. (c) Image on (b) after applying sigma filter for smoothing. (d) Image on (c) after applying gradient magnitude filter. (e) The final segment result after applying the level-set method.
was then applied to separate the 3-D vascular features inside and outside the targeted tumor. None of our cases had any posterior shadowing that would cause interference to assess vascular information. Vessel voxel processing and vessel tree construction In 3-D PDUS, the signal strength for blood vessels was included only in the red channel of the RGB vessel images. The value of blood vessels was between 0 and 255. A predefined threshold value, 120, was used to filter out vessels with relatively lower density of red blood cells (Chang et al. 2006, 2007). Thus, the binary vessel images of the main vessels were obtained (Fig. 3). Morphologic closing operators (Gonzalez and Woods 2002) were applied with two successive dilation expansions to fill out the cavities of vessel skeletons, followed by two successive erosions to smooth the vessel
Fig. 3. The vessel volume after thresholding (orange) and the vessel skeletons (red).
boundaries and restore the size. The 3-D thinning algorithm (Palagyi and Kuba 1998) was used to extract vessel skeletons from the main vessels to analyze vessel morphology and tortuosity (Fig. 3) (Chang et al. 2006, 2007). Each connected vessel skeleton was represented by a tree structure, called the vessel tree, which was builtfrom the root using a breadth first search (BFS) algorithm (Chang et al. 2006, 2007; Horowitz et al. 2007). The vessel skeletons inside and outside the tumors were used to construct the respective vessel trees. Extraction of vascular morphologic and tortuous features Vascular morphologic (VM) and vascular tortuous (VT) features obtained by recursively traversing vessel trees were used to analyze tumor vascularity inside and outside the tumor, respectively. The ‘‘significant vessel tree’’ was defined as the vessel tree with the largest sum of radii and hypothesized to be the most contributing vessel tree to tumor growth. To extract the most significant contributing vessel in each vessel tree, a ‘‘primary path’’" was defined as the path with the maximum sum of radii (Fig. 4). This primary path was traversed recursively to obtain features, including vessel radius and three tortuous features. Because the significant vessel tree dominated the main blood flow of the tumor, the vascular features were also extracted from the primary path of the significant vessel tree that was presumed to be the most important vessel skeleton. Seven VM features (Chang et al. 2007), including number of vessel trees (N), length of vessel trees (L), number of branches (B), radius of vessels (R), number of cycles (C), volume of vessel (VV) and vascularization index (VI), were computed, respectively, from inside and outside the vessel trees. Three VT features (Bullitt et al. 2003), including distance metric (DM), inflection count metric (ICM) and sum of angle metric (SOAM),
1862
Ultrasound in Medicine and Biology
Volume 38, Number 11, 2012
Table 1. Definitions of vascular morphologic and tortuous features on 3-D power Doppler breast ultrasound Abbreviations N0in and N0out in out and N10 N10 N5in and N5out out Lin all and Lall out Lin sig and Lsig
Bin and Bout out Rin mean and Rmean
Rout max
Fig. 4. Diagram for selecting the "primary path" within the "vessel tree." Each vessel tree had a primary path. In this case, the "primary path" was defined as the maximum sum of radii on pixels D, B, A, C, H and K, while the total sum of all radii of the "primary path" was 19 (61412121114). A–M represented labels of each node, while each number following A–M represented the radius of each node. A path, D-B-A-C-H-K, was defined as the ‘‘primary path’’ in this tree.
were extracted from the average assessment of primary paths (l) of all vessel trees and from the primary path of the significant vessel tree (Bullitt et al. 2004). The DM feature was denoted by the ratio between the path length of a bent vessel and the linear distance of two endpoints. For the ICM feature, the inflection point used to show the number of times that this primary path changed its orientation differentiated between two primary paths of vessels with the same DM feature but with different degree of tortuosity. The SOAM feature was used to analyze orientation changes voxel by voxel. If a vessel path changed orientation violently, the tortuosity of the vessel was large. Thus, the primary path of a vessel tree was traversed to sum up the orientation change of each voxel in the primary path and normalized by the length of the primary path. Thirty-three features were obtained from the sub-classification of VM and VT features. (Table 1) To simplify the multitude of all features, a wellknown classification method, the neural network machine (Duda et al. 2001; Hagan et al. 1996; Meinel et al. 2007), was used to extract more important features instead of manually diagnosing all features. To construct the neural network machine, some robust features were first selected from these 33 features using backward selection (Kohavi and John 1997). The multilayered perception (MLP) neural network (Duda et al. 2001; Hagan et al. 1996; Meinel et al. 2007) was used to identify benign and malignant tumors using the selected features and then evaluated by the k-fold cross-validation method (Kohavi and John 1997).
VI in and VI out VV in and VV out C out MDMin and MDMout DM_lin and DM_lout MICMin and MICMout ICM_lin and ICM_lout MSOAMin and MSOAMout SOAM_lin and SOAM_lout TV
Definition Total numbers of inside and outside vessel trees Vessel trees larger than ten voxels Vessel trees larger than five voxels, Total length of all inside and outside vessel trees Length of the significant vessel trees inside and outside vessel trees Total number of branches within inside and outside vessel trees Mean radius of primary path in each inside and outside vessel tree Maximum radius of primary paths in outside vessel trees Vascularization index inside and outside the tumor Volume of vessels inside and outside the tumor Number of cycle outside vessel trees Average distance metric computed from the primary path (l) in each inside and outside vessel trees Distance metric from the primary path (l) in the inside and outside significant vessel trees Average inflection count metric computed from primary path (l) in each inside and outside vessel trees Inflection count metric from the primary path (l) in the inside and outside significant vessel trees Average sum of angle metric computed from primary path (l) in each inside and outside vessel trees Sum of angle metric from the primary path (l) in the inside and outside significant vessel trees The tumor volume
Statistical analysis Group comparison of the VM and VT features between benign and malignant lesions was performed using the Student’s t-test (two-tailed) (Moore 2008). The performance of the selected vascular features from the inside/outside vessel trees, vascular features of the whole VOI (Chang et al. 2006), 3-D VI and tumor size was evaluated by the receiver operator characteristic (ROC) curve analysis (LABROC1, 1993; Charles E. Metz, MD, University of Chicago, Chicago, IL, USA) (Song 1997). The difference between the AZ values of the ROC curves was evaluated using the z-test. Diagnostic performance of binary logistic regression (Kimber et al. 1988) for predicting malignancy on 3-D power Doppler US images was estimated. The accuracy, sensitivity, specificity and positive and negative predictive values were obtained using the method with the best diagnostic performance. A p value of ,0.05 was regarded as statistically significant.
Vascular morphology and tortuosity analysis d Y.-C. CHANG et al.
Table 2. Comparison of 21 vascular morphologic (VM) features between benign and malignant breast mass lesions derived from 3-D power Doppler ultrasound Feature
Benign
Malignant
p value
N0in N0out in N10 out N10 N5in N5out Lin all ðmmÞ Lout all ðmmÞ Lin sig ðmmÞ Lout sig ðmmÞ Bin Bout Rin mean ðmmÞ Rout mean ðmmÞ Rout max ðmmÞ VI in VI out VV in ðmm3 Þ VV out ðmm3 Þ C out TVðmm3 Þ
0.03 6 0.26 3.70 6 4.52 ,0.01 0.75 6 1.39 0.02 6 0.13 1.60 6 2.51 0.05 6 4.17 10.70 6 19.81 0.03 6 0.27 4.57 6 7.68 ,0.01 0.5 6 1.14 ,0.01 0.14 6 0.21 0.36 6 0.32 0.001 6 0.007 0.002 6 0.004 0.27 6 1.47 34.57 6 67.23 ,0.01 137.63 6 130.39
0.45 6 1.07 9.02 6 4.52 0.02 6 0.14 2.06 6 1.26 0.15 6 0.46 4.08 6 2.40 0.85 6 2.25 67.14 6 73.16 0.58 6 1.46 42.04 6 62.97 ,0.01 4.42 6 6.29 0.008 6 0.06 0.67 6 0.65 1.43 6 1.55 0.02 6 0.08 0.01 6 0.04 19.24 6 85.26 827.20 6 3249.92 0.30 6 0.89 686.66 6 822.46
,0.01* ,0.01* 0.32 ,0.01* 0.04* ,0.01* ,0.01* ,0.01* ,0.01* ,0.01* 1 ,0.01* 0.32 ,0.01* ,0.01* 0.12 0.038* 0.11 0.08 0.02* ,0.01*
The Student’s t-test was used to compare each feature between benign and malignant lesion. A p value of ,0.05 was considered statistically significant (*).
RESULTS Vascular morphologic and tortuous features of 3-D power Doppler US Compared with benign lesions, malignant lesions had larger values of 21 VM features obtained from 3-D PDUS (Table 2). Malignant tumors were associated out with significantly larger vessel number (N0in ,N0out ,N10 , in out N5 and N5 ), longer length of total vessels and signifiout in out cant vessel trees (Lin all ,Lall ,Lsig and Lsig ) and more vessel out branches outside tumor (B ), larger values of the mean radius of primary paths outside tumor (Rout mean ), larger maximum radius of the primary paths in the outside vessel trees (Rout max ), larger vascularization index outside tumor (VI out ), more vascular cycles outside tumor (Cout ) and larger tumor volume (TV) (all p values of ,0. 05). Results of 12 VT features obtained from 3-D PDUS in benign and malignant breast lesions showed that all VT features of malignant lesions outside the tumor (MDMout, DM_lout, MICMout, ICM_lout, MSOAMout MSOAMout and SOAM_lout) were significantly larger than those of benign lesions (Table 3). In contrast, the values of several VT features inside tumor (MDMin, DM_lin, MICMin, ICM_lin) were higher in malignant lesions but none of the VT features was associated with significant difference between malignant and benign lesions. Compared with the VT features inside tumors, only VT features of outside tumor contour showed significant difference for differentiating malignant and benign lesions.
1863
Table 3. Comparison of 12 vascular tortuous (VT) features derived from 3-D power Doppler ultrasound between benign and malignant breast mass lesion Feature in
MDM MDMout DM_lin DM_lout MICMin MICMout ICM_lin ICM_lout MSOAMin MSOAMout SOAM_lin SOAM_lout
Benign
Malignant
p value
,0.01 0.47 6 0.63 ,0.01 0.47 6 0.64 ,0.01 0.80 6 1.78 ,0.01 0.87 6 1.90 ,0.01 0.17 6 0.42 ,0.01 0.19 6 0.45
0.02 6 0.18 1.66 6 1.51 0.02 6 0.18 1.75 6 1.65 0.02 6 0.18 6.35 6 13.64 0.02 6 0.18 7.05 6 14.79 ,0.01 0.57 6 0.51 ,0.01 0.65 6 0.57
0.32 ,0.01* 0.32 ,0.01* 0.32 ,0.01* 0.32 ,0.01* 1 ,0.01* 1 ,0.01*
The Student’s t-test was used to compare each feature between benign and malignant lesion. A p value of ,0.05 was considered statistically significant (*).
Results of input feature selection and performance of neural network After performing the back selection algorithm for all 33 features, 14 (11 VM features, 2 VT features and TV) were used as inputs of the neural network machine. Ten features were extracted from the vessel trees outside the tumor: total number of vessels (N0out ), number of braches (Bout ), number of cycles (C out ), total length of vessels out (Lout all ), length of the significant vessel tree (Lsig ), mean out radius of primary path (Rmean ), maximum radius of out primary paths (Rout max ), volume of vessels (VV ), mean out inflection count metric (MICM ) and inflection count metric of the significant vessel tree (ICM lout ). Furthermore, three features were computed from the vessel trees inside the tumor: total number of vessels (N0in ), number of vessels with larger than five voxels (N5in ) and volume of vessels (VV in ). The performance and number of misdiagnosed tumors were determined by applying the fivefold cross-validation method to the neural network machine (Tables 4 and 5). In the ROC curve analysis, the proposed system with 14 selected features (outside/inside VM-VT) showed the best performance, with Az value of 0.9188 significantly better than the that of using whole VM-VT (Az 5 0.8588) (p 5 0.0021), which used seven vascular features without using tumor contour segmentation (Chang et al. 2006), tumor volume (TV) (Az 5 0.7982) (p 5 0.0058) or 3-D VI (Az 5 0.6595) (Fig. 5). The accuracy, sensitivity, specificity, positive predictive value and negative predictive value of the proposed system using 14 neural network selected features were 89.38% (101/113), 84.91% (45/53), 93.33% (56/60), 91.84% (45/49) and 87.50% (56/64), respectively (Table 6). The accuracy, sensitivity and NPV of the 14 selected features with outside/inside VM-VT were significantly better than the other three
1864
Ultrasound in Medicine and Biology
Volume 38, Number 11, 2012
Table 4. Performance of the proposed method* with the neural network at different threshold values Diagnosed as False False malignant negative positive Threshold tumors Sensitivity Specificity (No.) (No.) 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1
36 38 43 48 49 51 53 55 58
0.6226 0.6603 0.7358 0.8302 0.8491 0.8492 0.8492 0.8492 0.8679
0.95 0.95 0.9333 0.9333 0.9333 0.9 0.8667 0.8333 0.8
20 18 14 9 8 8 8 8 7
3 3 4 4 4 6 8 10 12
Threshold was a cut-off value. If the malignancy probability evaluated from the neural network was larger than or equal to the threshold, the tumor was classified as malignant. * Using 13 vascular morphologic and tortuous features and tumor volume derived by 3-D power Doppler ultrasound.
methods using whole VM-VT, 3-D VI and tumor volume. These findings indicated the feasibility of using our proposed system as a screening tool to distinguish suspected lesions. Two correctly diagnosed benign and malignant cases, as well as misdiagnosed malignant cases, were also demonstrated (Figs. 6–9). DISCUSSION Compared with prior studies for quantitative analysis of 3-D PDUS that use global vascular morphologic (VM) and vascular tortuous (VT) features (Chang et al. 2006, 2007), the novelty of the present study is the addition of tumor contour information from the B-mode image for identifying VM and VT features outside and inside the tumor. The results show that the vascular features outside the tumor are useful for differentiating malignant from benign lesions. The findings suggest the importance of the peri-tumor vascular environment and the ‘‘significant vessel tree’’ of breast masses. Using combined VM and VT features inside and outside the tumor for discriminating benign from malignant breast masses provides the best performance (Az value 0.9188) compared with other methods (VM-VT of whole tumor, tumor volume and 3-D VI) (p , 0.05) (Fig. 5). Table 5. The number of misdiagnosed tumors using proposed system with 13 vascular morphologic and tortuous features and tumor volume at each test set of the fivefold cross-validation method at threshold 0.5 Test set
Malignant tumors
Benign tumors
1 2 3 4 5
1 of 10 1 of 10 3 of 11 2 of 11 1 of 11
0 of 12 1 of 12 1 of 12 1 of 12 1 of 12
Fig. 5. Comparison of receiver operator characteristic (ROC) curves computed using the proposed system with 14 selected features, including 13 vascular features outside or inside the segmented tumor contour and tumor volume (Outside/inside VM-VT; Az 0.9188), vascular system using seven vascular features without segmentation of the tumor contour (Whole VM-VT; Az 0.8588), simple system by tumor volume only (Az 0.7982) and conventional system with 3-D VI (Az 0.6595).
In earlier studies, quantification of tumor vascularity using VI, FI or VFI within the tumor obtained from 2-D or 3-D PDUS can correlate with MVD or disease prognosis (Chen et al. 2002; Cheng et al. 1999; Uzzan et al. 2004). Three-dimensional PDUS promises greater accuracy for its consistent sampling over the entire tumor without the problem of imaging plane variation of 2-D PDUS. Recent results use some quantification parameters from 3-D PDUS, such as VI, FI and VFI (Krestan et al. 2002), some VM features like vessel number, radius, branching, length and bifurcation and some VT features like distance metric (DM), inflection count metric (ICM) and sum of angle metric (SOAM) (Bullitt et al. 2003) for differentiating between benign and malignant breast masses. Regional variations of vessel distribution outside the tumor are different in benign and in malignant breast masses (Carson et al. 1997; Krestan et al. 2002). Shegal et al. measured vascularity of 2-D CD and PDUS with MVD correlation in 74 lesions and found gradient differences of tumor vascularity between malignant and benign breast masses (Carson et al. 1997). In their study, malignant masses exhibited strong gradient in vascularity while the benign masses had relatively uniform vascular distribution. Hsiao et al. (2008) used 3-D PDUS to measure vascularity within the tumor and shell around the tumor and found that malignancies had
Vascular morphology and tortuosity analysis d Y.-C. CHANG et al.
1865
Table 6. Comparison of the performance of using vascular morphology and vascular tortuosity features with and without tumor contour segmentation Item
Outside/inside VM-VT
Whole VM-VT
3-D VI
Tumor volume
Accuracy Sensitivity Specificity Positive predictive value (PPV) Negative predictive value (NPV)
89.38%*yz (101/113) 84.91%yz (45/53) 93.33% (56/60) 91.84% (45/49) 87.50%yz (56/64)
82.30%x (93/113) 81.13%xjj (43/53) 83.33%xjj (50/60) 81.13% (43/53) 83.33%xjj (50/60)
66.37% (75/113) 32.08% (17/53) 96.67% (58/60) 89.47% (17/19) 61.70% (58/94)
74.34% (84/113) 47.17% (25/53) 98.33% (59/60) 96.15% (25/26) 67.82% (59/87)
Outside/inside VM-VT 5 vascular morphology and tortuosity features of inside and outside vessel trees and tumor volume; Whole VM-VT 5 vascular morphologic and tortuous features of whole volume of interest (VOI); 3-D VI three-dimensional vascularization index (VI) of the whole tumor. Both methods were tested using the neural network machine based on a five-fold cross-validation method. Accuracy 5 (TP1TN)/(TP1TN1FP1FN) Sensitivity 5 TP/(TP1FN) Specificity 5 TN/(TN1FP) PPV 5 TP/(TP1FP) NPV 5 TN/(TN1FN) All p values were by c2 test. * p , 0.05 for comparison between ‘‘outside/inside VM VT’’ and ‘‘whole VM VT’’. y p , 0.05 for comparison between ‘‘outside/inside VM VT’’ and ‘‘3-D VI’’. z p , 0.05 for comparison between ‘‘outside/inside VM VT’’ and ‘‘tumor volume’’. x p , 0.05 for comparison between ‘‘whole VM VT’’ and ‘‘3-D VI’’. jj p , 0.05 for comparison between ‘‘whole VM VT’’ and ‘‘tumor volume’’.
higher VI, FI and VFI both within the tumor and in its shell (Krestan et al. 2002). Furthermore, the extent of peri-tumor vascular invasion in breast cancer is associated with unfavorable outcome (Colleoni et al. 2007).
Results of the current study suggest that not only intratumor vascularity but also the peri-tumor environment are important for tumor characterization (Tables 2 and 3). Tumors surrounded by non-neoplastic tissue with a high
Fig. 6. An example of true negative in a patient with benign fibrocystic change. The malignancy probability evaluated by the neural network was 0.06, N0in 5 0, N5in 5 0, VV in 5 0 mm3 , N0out 5 4, Bout 5 0, Cout 5 0, Lout all 5 6:40 mm, out out 3 out out Lout 5 3:50 mm, R 5 0:40 mm, VV 5 15:55 mm , MICM 5 1:15, ICM L 5 1:15 and TV 5 20.62 mm3. The sig mean values of these vascular parameters indicated that the benign tumor was associated with less vascularity and tortuosity. (a) Vessels (orange) after thresholding and its skeletons (red and white). (b) The significant vessel trees outside (red) and inside (white) the tumor. (c) The tumor contour (blue) and primary path in the significant vessel tree outside (red) and inside the tumor (white). (d) Original 2-D PDUS showed a small hypoechoic nodule (arrow) corresponding to the blue-color tumor in (c) on the right hand side and a vessel on the left hand side.
1866
Ultrasound in Medicine and Biology
Volume 38, Number 11, 2012
Fig. 7. An example of true positive in a patient with malignant invasive lobular carcinoma. The malignancy probability evaluated by the neural network was 0.99, N0in 5 5, N5in 5 1, VV in 5 24:78 mm3 , N0out 5 2, Bout 5 17, Cout 5 5, out out out 5 821:67 mm3 , MICM out 5 89.59, ICM Lout 5 95:67 Lout all 5 151:76 mm, Lsig 5 147:01 mm, Rmean 5 1:37 mm, VV and TV 51890.35 mm3. The values of these vascular parameters indicated that the malignant tumor was accompanied with more prominent vessel trees, more bending and twisting vessels. (a) Vessels (orange) after thresholding and its skeletons (red and white). (b) Significant vessel trees outside (red) and inside (white) the tumor. (c) The tumor contour (blue) and primary path in the significant vessel tree outside (red) and inside the tumor (white). (d) The original 2-D PDUS.
density of endothelial cell nitric oxide synthase (ecNOS)positive microvessels are associated with better recurrence-free and overall survival of breast cancer patients (Mortensen et al. 1999). It is assumed that the microvascular environment surrounding a tumor may be different between benign and malignant lesions and this may be useful for imaging analysis of breast tumors in future studies. In our study, the number, length, branch, radius and volume of vessels inside tumors were far lower than those outside tumors (Tables 2 and 3). Therefore, the intratumoral vascular information was less sufficient for analyzing VT features, which require more vascular information. In contrast, the VM features are the quantity evaluation of vessel distribution, which are easier to obtain. The malignant tumors had more vessels inside the tumor than benign tumors in both VM and VT features (Tables 2 and 3). In contrast to VM features, there were far fewer significant VT features inside tumors for differentiating benign and malignant lesions probably due to a smaller vascular size, length and diameter within the tumor. Tumor angiogenesis is crucial for tumor growth and metastases. The MVD has been demonstrated as
a potent prognostic indicator in breast cancer (Toi et al. 1995). Three-dimensional PDUS provides stronger appreciation of vascular morphology (Carson et al. 1997; Yang et al. 2002) and significantly higher correlation with MVD than 2-D PDUS (Yang et al. 2002). However, 3-D VI has the worst performance (accuracy 66.37%) for differentiating benign from malignant mass lesions. In contrast, accuracy and sensitivity increase significantly if the VM and VT features of the whole VOI are applied and become even higher if VM and VT features outside and inside the segmented tumor are used separately (Table 6), which corroborate the importance of the surrounding vascular environment. Prior studies have shown that malignant breast cancers usually have larger tumor sizes, more vessels and higher flow intensity than benign tumors (Hsiao et al. 2008; Sehgal et al. 2000; Yang et al. 2002), similar to the results of the present study. Histopathologic and immuno-histochemical analysis of angiogenesis indicate that both quantitative measures (i.e., microvessel density, total microvascular area, branching count and vascular diameter) and vascular morphology (e.g., shape, network and complexity) are relevant to prognosis of variable
Vascular morphology and tortuosity analysis d Y.-C. CHANG et al.
1867
Fig. 8. An example of false negative in a patient with invasive lobular carcinoma. The malignancy probability evaluated by neural network was 0.07, diagnosed benign using proposed system (Outside/inside VM-VT), N0in 5 0, out out N5in 5 0, VV in 5 1.29 mm3, N0out 5 13, Bout 5 4, Cout 5 0, Lout all 5 38:57 mm, Lsig 5 16:03 mm, Rmean 5 0:49 mm, VV out 5 484.48 mm3, MICM out 5 1.55, ICM lout 5 2.97 and TV 5 499.68 mm3. (a) Vessels (orange) after thresholding and its skeletons (red and white). (b) The significant vessel trees outside (red) and inside (white) the tumor. (c) The tumor contour (blue) and primary path in significant vessel tree outside (red) and inside the tumor (white). (d) The original 2-D PDUS.
tumors (Sharma et al. 2005). Measuring vascular tortuosity by magnetic resonance angiography can detect different types of cerebral abnormalities (Bullitt et al. 2003). Vessel tortuosity is considered a marker of treatment response in malignant gliomas (Bullitt, Ewend 2004). More interestingly, there is a significant difference in vascular density and vessel tortuosity in the transgenic mice model with or without myeloid cell-derived vascular endothelial growth factor (VEGF) (Weidemann et al. 2008). Thus, combined analysis of quantitative vasculature, vascular morphology and tortuosity via 3-D PDUS is important in approaching breast masses. Furthermore, the quantitative vascular features acquired by 3-D PDUS can potentially serve as imaging biomarkers, not only for the differentiation of benign from malignant lesions but also for the early prediction of tumor response to anti-angiogenesis therapy or neoadjuvant chemotherapy. Different methods of vessel partitioning are used to evaluate vascularity (Gokalp et al. 2009; LeCarpentier et al. 2008) with improvement of diagnostic performance. Simple qualitative and quantitative analysis of PDUS using the number of vessels in the lesion, distribution of tumor vessels, vessel morphology and spectral evaluation show an accuracy, sensitivity and specificity of 77.7%,
71.8% and 81.8%, respectively (Gokalp et al. 2009). Manual selection of region-of-interest (ROI) within the tumor lesion to build a polygon followed by computing simple vessel features (VI, FI, VFI) inside and outside the polygon showed accuracy, sensitivity and specificity of 81%, 94% and 69%, respectively (Huang et al. 2009). After applying the VM and VT features outside and inside tumor by 3-D PDUS in the present study, performance improved to 89.4% accuracy, 84.9% sensitivity and 93.3% specificity. There were some limitations to this study. There was no intraobserver validation for the same lesion and histopathologic correlation of the vascularity was lacking. In addition, this technique may be difficult to apply if the tumor is associated with prominent posterior shadowing that obscures the tumor vascularity. On the other hand, the proposed method is useful and verified on 113 tumors, including 60 benign and 53 malignant tumors. The dataset might be too small to completely prove the robustness of this diagnosis system. Therefore, a larger cohort study should be collected for robustness verification of this system in the future. In conclusion, separation of the vascular information inside and outside 3-D isosurface tumor contour obtained
1868
Ultrasound in Medicine and Biology
Volume 38, Number 11, 2012
Fig. 9. An example of false positive in a case of fibrocystic changes. The malignancy probability evaluated by the neural out network was 0.95, N0in 5 0, N5in 5 0, VV in 5 0 mm3, N0out 5 10, Bout 5 0, Cout 5 0, Lout all 5 45:54 mm, Lsig 5 18:84 mm, 3 3 out out out 5 0:55, mm VV 5 191.90 mm , MICM 5 3.37, ICM L 5 4.70 and TV 5 542.76 mm . (a) Vessels (orange) Rout mean after thresholding and its thinning skeletons (red and white). (b) Significant vessel trees outside (red) and inside (white) the tumor. (c) The tumor contour (blue) and primary path in the significant vessel tree outside (red) and inside the tumor (white). (d) The original 2-D PDUS.
from 3-D PDUS can provide more delicate quantitative measurement of VM and VT. This approach can significantly improve diagnostic accuracy in breast masses and has potential clinical value. Acknowledgments—The authors thank the National Science Council of the Republic of China for the financial support for this research (Contract No. NSC 96-2221-E-002-268-MY3).
REFERENCES Bullitt E, Ewend MG, Aylward S, Lin WL, Gerig G, Joshi S, Jung I, Muller K, Smith JK. Abnormal vessel tortuosity as a marker of treatment response of malignant gliomas: Preliminary report. Technol Cancer Res Treat 2004;3:577–584. Bullitt E, Gerig G, Pizer SM, Lin W, Aylward SR. Measuring tortuosity of the intracerebral vasculature from MRA images. IEEE Trans Med Imaging 2003;22:1163–1171. Bullitt E, Lin NU, Smith JK, Zeng D, Winer EP, Carey LA, Lin W, Ewend MG. Blood vessel morphologic changes depicted with MR angiography during treatment of brain metastases: A feasibility study. Radiology 2007;245:824–830. Carson PL, Moskalik AP, Govil A, Roubidoux MA, Fowlkes JB, Normolle D, Adler DD, Rubin JM, Helvie M. The 3D and 2D color flow display of breast masses. Ultrasound Med Biol 1997;23: 837–849. Chang RF, Huang SF, Moon WK, Lee YH, Chen DR. Computer algorithm for analysing breast tumor angiogenesis using 3-D power doppler ultrasound. Ultrasound Med Biol 2006;32:1499–1508. Chang RF, Huang SF, Moon WK, Lee YH, Chen DR. Solid breast masses: Neural network analysis of vascular features at three-
dimensional power Doppler US for benign or malignant classification. Radiology 2007;243:56–62. Chen CN, Cheng YM, Lin MT, Hsieh FJ, Lee PH, Chang KJ. Association of color Doppler vascularity index and microvessel density with survival in patients with gastric cancer. Ann Surg 2002;235: 512–518. Cheng WF, Lee CN, Chu JS, Chen CA, Chen TM, Shau WY, Hsieh CY, Hsieh FJ. Vascularity index as a novel parameter for the in vivo assessment of angiogenesis in patients with cervical carcinoma. Cancer 1999;85:651–657. Colleoni M, Rotmensz N, Maisonneuve P, Sonzogni A, Pruneri G, Casadio C, Luini A, Veronesi P, Intra M, Galimberti V, Torrisi R, Andrighetto S, Ghisini R, Goldhirsch A, Viale G. Prognostic role of the extent of peritumoral vascular invasion in operable breast cancer. Ann Oncol 2007;18:1632–1640. Duda RO, Hart PE, Stork DG. Pattern classification. New York: Wiley; 2001. Ferrara KW, Zagar B, Sokil-Melgar J, Algazi VR. High resolution 3D color flow mapping: Applied to the assessment of breast vasculature. Ultrasound Med Biol 1996;22:293–304. Gokalp G, Topal U, Kizilkaya E. Power Doppler sonography: Anything to add to BI-RADS US in solid breast masses? Eur J Radiol 2009;70: 77–85. Gonzalez RC, Woods RE. Digital image processing. Upper Saddle River, N.J: Prentice Hall; 2002. Hagan MT, Demuth HB, Beale MH. Neural network design. Boston: PWS Pub.; 1996. Horowitz E, Sahni S, Mehta DP. Fundamentals of data structures in C11. Summit, NJ: Silicon Press; 2007. Hsiao YH, Kuo SJ, Liang WM, Huang YL, Chen DR. Intra-tumor flow index can predict the malignant potential of breast tumor: Dependent on age and volume. Ultrasound Med Biol 2008;34:88–95.
Vascular morphology and tortuosity analysis d Y.-C. CHANG et al. Huang SF, Chang RF, Moon WK, Lee YH, Chen DR, Suri JS. Analysis of tumor vascularity using three-dimensional power Doppler ultrasound images. IEEE Trans Med Imaging 2008;27:320–330. Huang YL, Kuo SJ, Hsu CC, Tseng HS, Hsiao YH, Chen DR. Computeraided diagnosis for breast tumors by using vascularization of 3-D power Doppler ultrasound. Ultrasound Med Biol 2009;35: 1607–1614. Kimber A, Bowman AW, Cox D, Ross GJS, Gilchrist R, Davison AC, Ohagan A, Lawrance AJ, Pregibon D, Critchley F, Ekholm A, Titterington DM, Williams DA, Copas JB, Palmgren J. Binary regression-models for contaminated data—Discussion. J R Statist Soc Ser B Met 1988;50:253–265. Kohavi R, John GH. Wrappers for feature subset selection. Artif Intell 1997;97:273–324. Krestan CR, Riedl C, Memarsadeghi M, Rudas M, Pfarl G, Helbich TH. 3D-power Doppler ultrasound of breast lesions with a microbubble contrast agent. Acad Radiol 2002;9(Suppl 2):S384–S385. LeCarpentier GL, Roubidoux MA, Fowlkes JB, Krucker JF, Hunt KA, Paramagul C, Johnson TD, Thorson NJ, Engle KD, Carson PL. Suspicious breast lesions: Assessment of 3D doppler US indexes for classification in a test population and fourfold cross-validationscheme. Radiology 2008;249:463–470. Lee JS. Digital image smoothing and the sigma filter. Comput Vis Graph Image Proc 1983;24:255–269. Lee SW, Choi HY, Baek SY, Lim SM. Role of color and power Doppler imaging in differentiating between malignant and benign solid breast masses. J Clin Ultrasound 2002;30:459–464. Malladi R, Sethian JA, Vemuri BC. Shape modeling with front propagation - A level set approach. IEEE Trans Pattern Anal Mach Intell 1995;17:158–175. Meinel LA, Stolpen AH, Berbaum KS, Fajardo LL, Reinhardt JM. Breast MRI lesion classification: Improved performance of human readers with a backpropagation neural network computer-aided diagnosis (CAD) system. J Magn Reson Imaging 2007;25:89–95. Moore DS. The basic practice of statistics. New York: W.H. Freeman and Co.; 2008. Mortensen K, Holck S, Christensen IJ, Skouv J, Hougaard DM, Blom J, Larsson LI. Endothelial cell nitric oxide synthase in peritumoral microvessels is a favorable prognostic indicator in
1869
premenopausal breast cancer patients. Clin Cancer Res 1999;5: 1093–1097. Osher S, Sethian JA. Fronts propagating with curvature-dependent speed—Algorithms based on hamilton-jacobi formulations. J Comput Phys 1988;79:12–49. Palagyi K, Kuba AA. 3D 6-subiteration thinning algorithm for extracting medial lines. Pattern Recogn Lett 1998;19:613–627. Sehgal CM, Arger PH, Rowling SE, Conant EF, Reynolds C, Patton JA. Quantitative vascularity of breast masses by Doppler imaging: Regional variations and diagnostic implications. J Ultrasound Med 2000;19:427–440. Sharma S, Sharma MC, Sarkar C. Morphology of angiogenesis in human cancer: A conceptual overview, histoprognostic perspective and significance of neoangiogenesis. Histopathology 2005;46: 481–489. Song HH. Analysis of correlated ROC areas in diagnostic testing. Biometrics 1997;53:370–382. Strano S, Gombos EC, Friedland O, Mozes M. Color Doppler imaging of fibroadenomas of the breast with histopathologic correlation. J Clin Ultrasound 2004;32:317–322. Suri JS. Advances in diagnostic and therapeutic ultrasound imaging. Boston, London: Artech House; 2008. Toi M, Inada K, Suzuki H, Tominaga T. Tumor angiogenesis in breastcancer—Its importance as a prognostic indicator and the association with vascular endothelial growth-factor expression. Breast Cancer Res Treat 1995;36:193–204. Uzzan B, Nicolas P, Cucherat M. Perret Gr-Y. Microvessel density as a prognostic factor in women with breast cancer: A systematic review of the literature and meta-analysis. Cancer Res 2004;64: 2941–2955. Weidemann A, Stockmann C, Doedens A, Zhang N, Takeda N, Greenberg JI, Cheresh DA, Johnson RS. Deletion of vascular endothelial growth factor in myeloid cells accelerates tumorigenesis. Nature 2008;456:814–818. Yang WT, Tse GMK, Lam PKW, Metreweli C, Chang J. Correlation between color power Doppler sonographic measurement of breast tumor vasculature and immunohistochemical analysis of microvessel density for the quantitation of angiogenesis. J Ultrasound Med 2002;21:1227–1235.
http://dx.doi.org/10.1016/j.ultrasmedbio.2012.06.010
d
Original Contribution VASCULAR MORPHOLOGY AND TORTUOSITY ANALYSIS OF BREAST TUMOR INSIDE AND OUTSIDE CONTOUR BY 3-D POWER DOPPLER ULTRASOUND YEUN-CHUNG CHANG,* YAN-HAO HUANG,y CHIUN-SHENG HUANG,z and RUEY-FENG CHANGyx * Department of Medical Imaging, National Taiwan Univesity Hospital and National Taiwan University College of Medicine, Taipei, Taiwan; y Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan; z Department of Surgery, National Taiwan University Hospital and National Taiwan University College of Medicine, Taipei, Taiwan; and x Graduate Institute of Biomedical Electronics and Bioinformatics, National Taiwan University, Taipei, Taiwan (Received 20 December 2011; revised 16 June 2012; in final form 26 June 2012)
Abstract—This study aimed to evaluate morphologic and tortuous features of vessels inside and outside the tumor region on three-dimensional power Doppler ultrasonography (PDUS) in 113 breast mass lesions, including 60 benign and 53 malignant tumors. Compared with benign lesions, malignant breast lesions had significantly larger values of vascular morphologic and tortuous features and larger tumor sizes. The receiver operating characteristic curve analysis and Student’s t-test were used to estimate the performance of a proposed classification system using 13 vascular features and tumor size selected by the neural network. Accuracy, sensitivity, specificity, positive predictive value, negative predictive value and the AZ value of the diagnosis performance based on 14 features were 89.38% (101/113), 84.91% (45/53), 93.33% (56/60), 91.84% (45/49), 87.50% (56/64) and 0.9188, respectively. The three-dimensional PDUS morphologic and tortuous characteristics of blood vessels inside and outside breast mass lesions can be effectively used to classify benign and malignant tumors. (E-mail: [email protected]. edu.tw) Ó 2012 World Federation for Ultrasound in Medicine & Biology. Key Words: 3-D ultrasound, Breast tumor, Vascularity, Morphology, Tortuosity.
to differentiate between benign and malignant breast tumors (Hsiao et al. 2008; Sehgal et al. 2000). Vascular quantification using VI, flow index (FI) and vascularization flow index (VFI) from computer extracted volume-of-interest (VOI) and area outside the target lesion on 3-D PDUS can also differentiate between malignant and benign lesions (Huang et al. 2009). Moreover, vessel tortuosity, which is the estimation of the degree of vessel bending is considered to be related to intra-cranial malignancy and useful for assessing tumor activity (Bullitt et al. 2003, 2004, 2007). Compared with normal tissues, the vasculature of malignant lesion is more branching and tortuous and the vessels of malignant tumor are more chaotic and malformed than those of benign tumors (Huang et al. 2008). If the vessels are bent and twisted, the findings are usually associated with tumor malignancy (Chang et al. 2006, Chang et al. 2007). Combined morphologic and tortuous vascular features from 3-D PDUS have higher accuracy in differentiating benign from malignant masses than using 3-D VI alone (Chang et al. 2006, 2007). Good correlation of anatomic vascular distribution between 3-D PDUS and histopathology has also been reported (Strano et al. 2004). Doppler features suggestive
INTRODUCTION Angiogenesis in breast cancer is commonly assessed and quantified using microvessel density (MVD) (Uzzan et al. 2004). Color Doppler ultrasound vascularization index (VI), defined as the number of vessel pixels divided by the number of all pixels, reportedly correlates with MVD (Chen et al. 1999, 2002). Since vascular quantification from two-dimensional (2-D) ultrasound (US) may vary due to difference of selected planes, VI obtained from three-dimensional (3-D) power Doppler ultrasound (PDUS) is more reliable and reproducible for its capacity for volumetric coverage and global assessment of tumor vascularity (Ferrara et al. 1996; Krestan et al. 2002). Compared with 2-D US, 3-D color flow display provides stronger subjective appreciation of vascular morphology (Carson et al. 1997) and significantly higher correlation with MVD (Yang et al. 2002). Recent studies using 3-D PDUS show that vascularity surrounding the tumor and intra-tumor vascularity on power Doppler US are useful Address correspondence to: Ruey-Feng Chang, PhD, Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan 106, R.O.C. E-mail: [email protected] 1859
1860
Ultrasound in Medicine and Biology
Volume 38, Number 11, 2012
of malignant lesions include the presence of both peripheral and central vascularity, as well as penetrating and branching vessels (Lee et al. 2002; Strano et al. 2004). Analysis of vascular morphology and distribution is important to obtain spatial information of tumor heterogeneity and to differentiate breast tumors (Gokalp et al. 2009; LeCarpentier et al. 2008). The purpose of the current study was to assess the diagnostic performance of combined 3-D vascular morphologic and tortuous features in comparison to previous methods for discriminating benign from malignant tumors. The grey-scale tumor margin was used to categorize vascular distribution inside and outside the targeted mass, with quantification of vascular morphologic and tortuous features. MATERIALS AND METHODS Patients One hundred and thirteen (113) consecutive solid tumors (range, 0.2–4.0 cm; mean, 1.8 6 0.9 cm) that received 3-D PDUS before undergoing surgical excision or percutaneous needle biopsy due to suspicious US findings between June 2008 and June 2009 were evaluated after excluding nine patients (7%) with a tumor size larger than 4 cm, which could not be included in one 3-D acquisition. All benign lesions, histopathologically confirmed, were followed-up for stability for at least 2 years. Sixty were benign lesions, including nine fibroadenomas, 39 focal fibrocystic changes, five papillomas and seven nonspecified fibro-epithelial lesions. Fiftythree were malignant, including 34 invasive ductal carcinomas (IDC), 11 ductal carcinoma in situ (DCIS) and eight invasive lobular carcinomas. Informed consent was obtained from each patient prior to performing a biopsy. Institutional review board (IRB) approval was obtained for this study and informed consent was waived for this retrospective review. Ultrasound imaging All US images were acquired using a 3-D power Doppler US scanner (Voluson 730; GE Healthcare, Zipf, Austria) and 6–12 MHz dedicated volume transducer by three radiologists with 3–8 years of experience in breast ultrasound. A suitable volume box size of 4 3 4 cm2 with adjustable depth was used to include the lesion with appropriate surrounding tissue and vasculature and the sweep angle was 25 –30 . The 3-D volume files were then saved in Cartesian coordinates using the 4-DView program with the US scanner. Lesions larger than 4.0 cm were excluded. Imaging preprocessing and tumor segmentation In 3-D power Doppler data, the vessel image, which contains the information about the density of red blood
Fig. 1. Flow chart of the proposed computer-aided diagnosis (CAD) system.
cells and grey B-mode image were obtained simultaneously and saved in the same image file. Then, the vessel image and grey B-mode image were further divided into two separate image files to extract vessels and tumor information, respectively. The tumor region was segmented from B-mode grey-scale ultrasound images. Three 3-D image processing methods were adopted to reduce noises and enhance the contrast for tumor segmentation. A 3-D sigmoid filter (Suri 2008) was used to enhance the gaps of intensity distributions between objects, while the sigma edge-preserving filter (Lee 1983) was applied to smooth images and preserve the boundary information (Fig. 1). Finally, a gradient magnitude filter (Gonzalez and Woods 2002) was used to extract the input data of the segmentation method (Fig. 2). The tumor extraction algorithm was based on the 3-D level set method by which a stable contour was obtained after manually selecting a seed within tumor and iteratively evolving the isosurface (Malladi et al. 1995; Osher and Sethian 1988). The user reviewed all the 2-D slices extracted from the 3-D B-mode image. The tumor center on the slice containing the maximum tumor contour or dimension was selected as the seed. Additional seeds on other slices could be manually added for more precise contour if the tumor was not well segmented. In this study, the average number of seeds was 1.3 (range, 1–5 points). The segmented tumor region
Vascular morphology and tortuosity analysis d Y.-C. CHANG et al.
1861
Fig. 2. The process of segmentation of grey-scale ultrasound image. (a) Original grey-scale image. (b) Image on (a) after applying sigmoid filter. (c) Image on (b) after applying sigma filter for smoothing. (d) Image on (c) after applying gradient magnitude filter. (e) The final segment result after applying the level-set method.
was then applied to separate the 3-D vascular features inside and outside the targeted tumor. None of our cases had any posterior shadowing that would cause interference to assess vascular information. Vessel voxel processing and vessel tree construction In 3-D PDUS, the signal strength for blood vessels was included only in the red channel of the RGB vessel images. The value of blood vessels was between 0 and 255. A predefined threshold value, 120, was used to filter out vessels with relatively lower density of red blood cells (Chang et al. 2006, 2007). Thus, the binary vessel images of the main vessels were obtained (Fig. 3). Morphologic closing operators (Gonzalez and Woods 2002) were applied with two successive dilation expansions to fill out the cavities of vessel skeletons, followed by two successive erosions to smooth the vessel
Fig. 3. The vessel volume after thresholding (orange) and the vessel skeletons (red).
boundaries and restore the size. The 3-D thinning algorithm (Palagyi and Kuba 1998) was used to extract vessel skeletons from the main vessels to analyze vessel morphology and tortuosity (Fig. 3) (Chang et al. 2006, 2007). Each connected vessel skeleton was represented by a tree structure, called the vessel tree, which was builtfrom the root using a breadth first search (BFS) algorithm (Chang et al. 2006, 2007; Horowitz et al. 2007). The vessel skeletons inside and outside the tumors were used to construct the respective vessel trees. Extraction of vascular morphologic and tortuous features Vascular morphologic (VM) and vascular tortuous (VT) features obtained by recursively traversing vessel trees were used to analyze tumor vascularity inside and outside the tumor, respectively. The ‘‘significant vessel tree’’ was defined as the vessel tree with the largest sum of radii and hypothesized to be the most contributing vessel tree to tumor growth. To extract the most significant contributing vessel in each vessel tree, a ‘‘primary path’’" was defined as the path with the maximum sum of radii (Fig. 4). This primary path was traversed recursively to obtain features, including vessel radius and three tortuous features. Because the significant vessel tree dominated the main blood flow of the tumor, the vascular features were also extracted from the primary path of the significant vessel tree that was presumed to be the most important vessel skeleton. Seven VM features (Chang et al. 2007), including number of vessel trees (N), length of vessel trees (L), number of branches (B), radius of vessels (R), number of cycles (C), volume of vessel (VV) and vascularization index (VI), were computed, respectively, from inside and outside the vessel trees. Three VT features (Bullitt et al. 2003), including distance metric (DM), inflection count metric (ICM) and sum of angle metric (SOAM),
1862
Ultrasound in Medicine and Biology
Volume 38, Number 11, 2012
Table 1. Definitions of vascular morphologic and tortuous features on 3-D power Doppler breast ultrasound Abbreviations N0in and N0out in out and N10 N10 N5in and N5out out Lin all and Lall out Lin sig and Lsig
Bin and Bout out Rin mean and Rmean
Rout max
Fig. 4. Diagram for selecting the "primary path" within the "vessel tree." Each vessel tree had a primary path. In this case, the "primary path" was defined as the maximum sum of radii on pixels D, B, A, C, H and K, while the total sum of all radii of the "primary path" was 19 (61412121114). A–M represented labels of each node, while each number following A–M represented the radius of each node. A path, D-B-A-C-H-K, was defined as the ‘‘primary path’’ in this tree.
were extracted from the average assessment of primary paths (l) of all vessel trees and from the primary path of the significant vessel tree (Bullitt et al. 2004). The DM feature was denoted by the ratio between the path length of a bent vessel and the linear distance of two endpoints. For the ICM feature, the inflection point used to show the number of times that this primary path changed its orientation differentiated between two primary paths of vessels with the same DM feature but with different degree of tortuosity. The SOAM feature was used to analyze orientation changes voxel by voxel. If a vessel path changed orientation violently, the tortuosity of the vessel was large. Thus, the primary path of a vessel tree was traversed to sum up the orientation change of each voxel in the primary path and normalized by the length of the primary path. Thirty-three features were obtained from the sub-classification of VM and VT features. (Table 1) To simplify the multitude of all features, a wellknown classification method, the neural network machine (Duda et al. 2001; Hagan et al. 1996; Meinel et al. 2007), was used to extract more important features instead of manually diagnosing all features. To construct the neural network machine, some robust features were first selected from these 33 features using backward selection (Kohavi and John 1997). The multilayered perception (MLP) neural network (Duda et al. 2001; Hagan et al. 1996; Meinel et al. 2007) was used to identify benign and malignant tumors using the selected features and then evaluated by the k-fold cross-validation method (Kohavi and John 1997).
VI in and VI out VV in and VV out C out MDMin and MDMout DM_lin and DM_lout MICMin and MICMout ICM_lin and ICM_lout MSOAMin and MSOAMout SOAM_lin and SOAM_lout TV
Definition Total numbers of inside and outside vessel trees Vessel trees larger than ten voxels Vessel trees larger than five voxels, Total length of all inside and outside vessel trees Length of the significant vessel trees inside and outside vessel trees Total number of branches within inside and outside vessel trees Mean radius of primary path in each inside and outside vessel tree Maximum radius of primary paths in outside vessel trees Vascularization index inside and outside the tumor Volume of vessels inside and outside the tumor Number of cycle outside vessel trees Average distance metric computed from the primary path (l) in each inside and outside vessel trees Distance metric from the primary path (l) in the inside and outside significant vessel trees Average inflection count metric computed from primary path (l) in each inside and outside vessel trees Inflection count metric from the primary path (l) in the inside and outside significant vessel trees Average sum of angle metric computed from primary path (l) in each inside and outside vessel trees Sum of angle metric from the primary path (l) in the inside and outside significant vessel trees The tumor volume
Statistical analysis Group comparison of the VM and VT features between benign and malignant lesions was performed using the Student’s t-test (two-tailed) (Moore 2008). The performance of the selected vascular features from the inside/outside vessel trees, vascular features of the whole VOI (Chang et al. 2006), 3-D VI and tumor size was evaluated by the receiver operator characteristic (ROC) curve analysis (LABROC1, 1993; Charles E. Metz, MD, University of Chicago, Chicago, IL, USA) (Song 1997). The difference between the AZ values of the ROC curves was evaluated using the z-test. Diagnostic performance of binary logistic regression (Kimber et al. 1988) for predicting malignancy on 3-D power Doppler US images was estimated. The accuracy, sensitivity, specificity and positive and negative predictive values were obtained using the method with the best diagnostic performance. A p value of ,0.05 was regarded as statistically significant.
Vascular morphology and tortuosity analysis d Y.-C. CHANG et al.
Table 2. Comparison of 21 vascular morphologic (VM) features between benign and malignant breast mass lesions derived from 3-D power Doppler ultrasound Feature
Benign
Malignant
p value
N0in N0out in N10 out N10 N5in N5out Lin all ðmmÞ Lout all ðmmÞ Lin sig ðmmÞ Lout sig ðmmÞ Bin Bout Rin mean ðmmÞ Rout mean ðmmÞ Rout max ðmmÞ VI in VI out VV in ðmm3 Þ VV out ðmm3 Þ C out TVðmm3 Þ
0.03 6 0.26 3.70 6 4.52 ,0.01 0.75 6 1.39 0.02 6 0.13 1.60 6 2.51 0.05 6 4.17 10.70 6 19.81 0.03 6 0.27 4.57 6 7.68 ,0.01 0.5 6 1.14 ,0.01 0.14 6 0.21 0.36 6 0.32 0.001 6 0.007 0.002 6 0.004 0.27 6 1.47 34.57 6 67.23 ,0.01 137.63 6 130.39
0.45 6 1.07 9.02 6 4.52 0.02 6 0.14 2.06 6 1.26 0.15 6 0.46 4.08 6 2.40 0.85 6 2.25 67.14 6 73.16 0.58 6 1.46 42.04 6 62.97 ,0.01 4.42 6 6.29 0.008 6 0.06 0.67 6 0.65 1.43 6 1.55 0.02 6 0.08 0.01 6 0.04 19.24 6 85.26 827.20 6 3249.92 0.30 6 0.89 686.66 6 822.46
,0.01* ,0.01* 0.32 ,0.01* 0.04* ,0.01* ,0.01* ,0.01* ,0.01* ,0.01* 1 ,0.01* 0.32 ,0.01* ,0.01* 0.12 0.038* 0.11 0.08 0.02* ,0.01*
The Student’s t-test was used to compare each feature between benign and malignant lesion. A p value of ,0.05 was considered statistically significant (*).
RESULTS Vascular morphologic and tortuous features of 3-D power Doppler US Compared with benign lesions, malignant lesions had larger values of 21 VM features obtained from 3-D PDUS (Table 2). Malignant tumors were associated out with significantly larger vessel number (N0in ,N0out ,N10 , in out N5 and N5 ), longer length of total vessels and signifiout in out cant vessel trees (Lin all ,Lall ,Lsig and Lsig ) and more vessel out branches outside tumor (B ), larger values of the mean radius of primary paths outside tumor (Rout mean ), larger maximum radius of the primary paths in the outside vessel trees (Rout max ), larger vascularization index outside tumor (VI out ), more vascular cycles outside tumor (Cout ) and larger tumor volume (TV) (all p values of ,0. 05). Results of 12 VT features obtained from 3-D PDUS in benign and malignant breast lesions showed that all VT features of malignant lesions outside the tumor (MDMout, DM_lout, MICMout, ICM_lout, MSOAMout MSOAMout and SOAM_lout) were significantly larger than those of benign lesions (Table 3). In contrast, the values of several VT features inside tumor (MDMin, DM_lin, MICMin, ICM_lin) were higher in malignant lesions but none of the VT features was associated with significant difference between malignant and benign lesions. Compared with the VT features inside tumors, only VT features of outside tumor contour showed significant difference for differentiating malignant and benign lesions.
1863
Table 3. Comparison of 12 vascular tortuous (VT) features derived from 3-D power Doppler ultrasound between benign and malignant breast mass lesion Feature in
MDM MDMout DM_lin DM_lout MICMin MICMout ICM_lin ICM_lout MSOAMin MSOAMout SOAM_lin SOAM_lout
Benign
Malignant
p value
,0.01 0.47 6 0.63 ,0.01 0.47 6 0.64 ,0.01 0.80 6 1.78 ,0.01 0.87 6 1.90 ,0.01 0.17 6 0.42 ,0.01 0.19 6 0.45
0.02 6 0.18 1.66 6 1.51 0.02 6 0.18 1.75 6 1.65 0.02 6 0.18 6.35 6 13.64 0.02 6 0.18 7.05 6 14.79 ,0.01 0.57 6 0.51 ,0.01 0.65 6 0.57
0.32 ,0.01* 0.32 ,0.01* 0.32 ,0.01* 0.32 ,0.01* 1 ,0.01* 1 ,0.01*
The Student’s t-test was used to compare each feature between benign and malignant lesion. A p value of ,0.05 was considered statistically significant (*).
Results of input feature selection and performance of neural network After performing the back selection algorithm for all 33 features, 14 (11 VM features, 2 VT features and TV) were used as inputs of the neural network machine. Ten features were extracted from the vessel trees outside the tumor: total number of vessels (N0out ), number of braches (Bout ), number of cycles (C out ), total length of vessels out (Lout all ), length of the significant vessel tree (Lsig ), mean out radius of primary path (Rmean ), maximum radius of out primary paths (Rout max ), volume of vessels (VV ), mean out inflection count metric (MICM ) and inflection count metric of the significant vessel tree (ICM lout ). Furthermore, three features were computed from the vessel trees inside the tumor: total number of vessels (N0in ), number of vessels with larger than five voxels (N5in ) and volume of vessels (VV in ). The performance and number of misdiagnosed tumors were determined by applying the fivefold cross-validation method to the neural network machine (Tables 4 and 5). In the ROC curve analysis, the proposed system with 14 selected features (outside/inside VM-VT) showed the best performance, with Az value of 0.9188 significantly better than the that of using whole VM-VT (Az 5 0.8588) (p 5 0.0021), which used seven vascular features without using tumor contour segmentation (Chang et al. 2006), tumor volume (TV) (Az 5 0.7982) (p 5 0.0058) or 3-D VI (Az 5 0.6595) (Fig. 5). The accuracy, sensitivity, specificity, positive predictive value and negative predictive value of the proposed system using 14 neural network selected features were 89.38% (101/113), 84.91% (45/53), 93.33% (56/60), 91.84% (45/49) and 87.50% (56/64), respectively (Table 6). The accuracy, sensitivity and NPV of the 14 selected features with outside/inside VM-VT were significantly better than the other three
1864
Ultrasound in Medicine and Biology
Volume 38, Number 11, 2012
Table 4. Performance of the proposed method* with the neural network at different threshold values Diagnosed as False False malignant negative positive Threshold tumors Sensitivity Specificity (No.) (No.) 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1
36 38 43 48 49 51 53 55 58
0.6226 0.6603 0.7358 0.8302 0.8491 0.8492 0.8492 0.8492 0.8679
0.95 0.95 0.9333 0.9333 0.9333 0.9 0.8667 0.8333 0.8
20 18 14 9 8 8 8 8 7
3 3 4 4 4 6 8 10 12
Threshold was a cut-off value. If the malignancy probability evaluated from the neural network was larger than or equal to the threshold, the tumor was classified as malignant. * Using 13 vascular morphologic and tortuous features and tumor volume derived by 3-D power Doppler ultrasound.
methods using whole VM-VT, 3-D VI and tumor volume. These findings indicated the feasibility of using our proposed system as a screening tool to distinguish suspected lesions. Two correctly diagnosed benign and malignant cases, as well as misdiagnosed malignant cases, were also demonstrated (Figs. 6–9). DISCUSSION Compared with prior studies for quantitative analysis of 3-D PDUS that use global vascular morphologic (VM) and vascular tortuous (VT) features (Chang et al. 2006, 2007), the novelty of the present study is the addition of tumor contour information from the B-mode image for identifying VM and VT features outside and inside the tumor. The results show that the vascular features outside the tumor are useful for differentiating malignant from benign lesions. The findings suggest the importance of the peri-tumor vascular environment and the ‘‘significant vessel tree’’ of breast masses. Using combined VM and VT features inside and outside the tumor for discriminating benign from malignant breast masses provides the best performance (Az value 0.9188) compared with other methods (VM-VT of whole tumor, tumor volume and 3-D VI) (p , 0.05) (Fig. 5). Table 5. The number of misdiagnosed tumors using proposed system with 13 vascular morphologic and tortuous features and tumor volume at each test set of the fivefold cross-validation method at threshold 0.5 Test set
Malignant tumors
Benign tumors
1 2 3 4 5
1 of 10 1 of 10 3 of 11 2 of 11 1 of 11
0 of 12 1 of 12 1 of 12 1 of 12 1 of 12
Fig. 5. Comparison of receiver operator characteristic (ROC) curves computed using the proposed system with 14 selected features, including 13 vascular features outside or inside the segmented tumor contour and tumor volume (Outside/inside VM-VT; Az 0.9188), vascular system using seven vascular features without segmentation of the tumor contour (Whole VM-VT; Az 0.8588), simple system by tumor volume only (Az 0.7982) and conventional system with 3-D VI (Az 0.6595).
In earlier studies, quantification of tumor vascularity using VI, FI or VFI within the tumor obtained from 2-D or 3-D PDUS can correlate with MVD or disease prognosis (Chen et al. 2002; Cheng et al. 1999; Uzzan et al. 2004). Three-dimensional PDUS promises greater accuracy for its consistent sampling over the entire tumor without the problem of imaging plane variation of 2-D PDUS. Recent results use some quantification parameters from 3-D PDUS, such as VI, FI and VFI (Krestan et al. 2002), some VM features like vessel number, radius, branching, length and bifurcation and some VT features like distance metric (DM), inflection count metric (ICM) and sum of angle metric (SOAM) (Bullitt et al. 2003) for differentiating between benign and malignant breast masses. Regional variations of vessel distribution outside the tumor are different in benign and in malignant breast masses (Carson et al. 1997; Krestan et al. 2002). Shegal et al. measured vascularity of 2-D CD and PDUS with MVD correlation in 74 lesions and found gradient differences of tumor vascularity between malignant and benign breast masses (Carson et al. 1997). In their study, malignant masses exhibited strong gradient in vascularity while the benign masses had relatively uniform vascular distribution. Hsiao et al. (2008) used 3-D PDUS to measure vascularity within the tumor and shell around the tumor and found that malignancies had
Vascular morphology and tortuosity analysis d Y.-C. CHANG et al.
1865
Table 6. Comparison of the performance of using vascular morphology and vascular tortuosity features with and without tumor contour segmentation Item
Outside/inside VM-VT
Whole VM-VT
3-D VI
Tumor volume
Accuracy Sensitivity Specificity Positive predictive value (PPV) Negative predictive value (NPV)
89.38%*yz (101/113) 84.91%yz (45/53) 93.33% (56/60) 91.84% (45/49) 87.50%yz (56/64)
82.30%x (93/113) 81.13%xjj (43/53) 83.33%xjj (50/60) 81.13% (43/53) 83.33%xjj (50/60)
66.37% (75/113) 32.08% (17/53) 96.67% (58/60) 89.47% (17/19) 61.70% (58/94)
74.34% (84/113) 47.17% (25/53) 98.33% (59/60) 96.15% (25/26) 67.82% (59/87)
Outside/inside VM-VT 5 vascular morphology and tortuosity features of inside and outside vessel trees and tumor volume; Whole VM-VT 5 vascular morphologic and tortuous features of whole volume of interest (VOI); 3-D VI three-dimensional vascularization index (VI) of the whole tumor. Both methods were tested using the neural network machine based on a five-fold cross-validation method. Accuracy 5 (TP1TN)/(TP1TN1FP1FN) Sensitivity 5 TP/(TP1FN) Specificity 5 TN/(TN1FP) PPV 5 TP/(TP1FP) NPV 5 TN/(TN1FN) All p values were by c2 test. * p , 0.05 for comparison between ‘‘outside/inside VM VT’’ and ‘‘whole VM VT’’. y p , 0.05 for comparison between ‘‘outside/inside VM VT’’ and ‘‘3-D VI’’. z p , 0.05 for comparison between ‘‘outside/inside VM VT’’ and ‘‘tumor volume’’. x p , 0.05 for comparison between ‘‘whole VM VT’’ and ‘‘3-D VI’’. jj p , 0.05 for comparison between ‘‘whole VM VT’’ and ‘‘tumor volume’’.
higher VI, FI and VFI both within the tumor and in its shell (Krestan et al. 2002). Furthermore, the extent of peri-tumor vascular invasion in breast cancer is associated with unfavorable outcome (Colleoni et al. 2007).
Results of the current study suggest that not only intratumor vascularity but also the peri-tumor environment are important for tumor characterization (Tables 2 and 3). Tumors surrounded by non-neoplastic tissue with a high
Fig. 6. An example of true negative in a patient with benign fibrocystic change. The malignancy probability evaluated by the neural network was 0.06, N0in 5 0, N5in 5 0, VV in 5 0 mm3 , N0out 5 4, Bout 5 0, Cout 5 0, Lout all 5 6:40 mm, out out 3 out out Lout 5 3:50 mm, R 5 0:40 mm, VV 5 15:55 mm , MICM 5 1:15, ICM L 5 1:15 and TV 5 20.62 mm3. The sig mean values of these vascular parameters indicated that the benign tumor was associated with less vascularity and tortuosity. (a) Vessels (orange) after thresholding and its skeletons (red and white). (b) The significant vessel trees outside (red) and inside (white) the tumor. (c) The tumor contour (blue) and primary path in the significant vessel tree outside (red) and inside the tumor (white). (d) Original 2-D PDUS showed a small hypoechoic nodule (arrow) corresponding to the blue-color tumor in (c) on the right hand side and a vessel on the left hand side.
1866
Ultrasound in Medicine and Biology
Volume 38, Number 11, 2012
Fig. 7. An example of true positive in a patient with malignant invasive lobular carcinoma. The malignancy probability evaluated by the neural network was 0.99, N0in 5 5, N5in 5 1, VV in 5 24:78 mm3 , N0out 5 2, Bout 5 17, Cout 5 5, out out out 5 821:67 mm3 , MICM out 5 89.59, ICM Lout 5 95:67 Lout all 5 151:76 mm, Lsig 5 147:01 mm, Rmean 5 1:37 mm, VV and TV 51890.35 mm3. The values of these vascular parameters indicated that the malignant tumor was accompanied with more prominent vessel trees, more bending and twisting vessels. (a) Vessels (orange) after thresholding and its skeletons (red and white). (b) Significant vessel trees outside (red) and inside (white) the tumor. (c) The tumor contour (blue) and primary path in the significant vessel tree outside (red) and inside the tumor (white). (d) The original 2-D PDUS.
density of endothelial cell nitric oxide synthase (ecNOS)positive microvessels are associated with better recurrence-free and overall survival of breast cancer patients (Mortensen et al. 1999). It is assumed that the microvascular environment surrounding a tumor may be different between benign and malignant lesions and this may be useful for imaging analysis of breast tumors in future studies. In our study, the number, length, branch, radius and volume of vessels inside tumors were far lower than those outside tumors (Tables 2 and 3). Therefore, the intratumoral vascular information was less sufficient for analyzing VT features, which require more vascular information. In contrast, the VM features are the quantity evaluation of vessel distribution, which are easier to obtain. The malignant tumors had more vessels inside the tumor than benign tumors in both VM and VT features (Tables 2 and 3). In contrast to VM features, there were far fewer significant VT features inside tumors for differentiating benign and malignant lesions probably due to a smaller vascular size, length and diameter within the tumor. Tumor angiogenesis is crucial for tumor growth and metastases. The MVD has been demonstrated as
a potent prognostic indicator in breast cancer (Toi et al. 1995). Three-dimensional PDUS provides stronger appreciation of vascular morphology (Carson et al. 1997; Yang et al. 2002) and significantly higher correlation with MVD than 2-D PDUS (Yang et al. 2002). However, 3-D VI has the worst performance (accuracy 66.37%) for differentiating benign from malignant mass lesions. In contrast, accuracy and sensitivity increase significantly if the VM and VT features of the whole VOI are applied and become even higher if VM and VT features outside and inside the segmented tumor are used separately (Table 6), which corroborate the importance of the surrounding vascular environment. Prior studies have shown that malignant breast cancers usually have larger tumor sizes, more vessels and higher flow intensity than benign tumors (Hsiao et al. 2008; Sehgal et al. 2000; Yang et al. 2002), similar to the results of the present study. Histopathologic and immuno-histochemical analysis of angiogenesis indicate that both quantitative measures (i.e., microvessel density, total microvascular area, branching count and vascular diameter) and vascular morphology (e.g., shape, network and complexity) are relevant to prognosis of variable
Vascular morphology and tortuosity analysis d Y.-C. CHANG et al.
1867
Fig. 8. An example of false negative in a patient with invasive lobular carcinoma. The malignancy probability evaluated by neural network was 0.07, diagnosed benign using proposed system (Outside/inside VM-VT), N0in 5 0, out out N5in 5 0, VV in 5 1.29 mm3, N0out 5 13, Bout 5 4, Cout 5 0, Lout all 5 38:57 mm, Lsig 5 16:03 mm, Rmean 5 0:49 mm, VV out 5 484.48 mm3, MICM out 5 1.55, ICM lout 5 2.97 and TV 5 499.68 mm3. (a) Vessels (orange) after thresholding and its skeletons (red and white). (b) The significant vessel trees outside (red) and inside (white) the tumor. (c) The tumor contour (blue) and primary path in significant vessel tree outside (red) and inside the tumor (white). (d) The original 2-D PDUS.
tumors (Sharma et al. 2005). Measuring vascular tortuosity by magnetic resonance angiography can detect different types of cerebral abnormalities (Bullitt et al. 2003). Vessel tortuosity is considered a marker of treatment response in malignant gliomas (Bullitt, Ewend 2004). More interestingly, there is a significant difference in vascular density and vessel tortuosity in the transgenic mice model with or without myeloid cell-derived vascular endothelial growth factor (VEGF) (Weidemann et al. 2008). Thus, combined analysis of quantitative vasculature, vascular morphology and tortuosity via 3-D PDUS is important in approaching breast masses. Furthermore, the quantitative vascular features acquired by 3-D PDUS can potentially serve as imaging biomarkers, not only for the differentiation of benign from malignant lesions but also for the early prediction of tumor response to anti-angiogenesis therapy or neoadjuvant chemotherapy. Different methods of vessel partitioning are used to evaluate vascularity (Gokalp et al. 2009; LeCarpentier et al. 2008) with improvement of diagnostic performance. Simple qualitative and quantitative analysis of PDUS using the number of vessels in the lesion, distribution of tumor vessels, vessel morphology and spectral evaluation show an accuracy, sensitivity and specificity of 77.7%,
71.8% and 81.8%, respectively (Gokalp et al. 2009). Manual selection of region-of-interest (ROI) within the tumor lesion to build a polygon followed by computing simple vessel features (VI, FI, VFI) inside and outside the polygon showed accuracy, sensitivity and specificity of 81%, 94% and 69%, respectively (Huang et al. 2009). After applying the VM and VT features outside and inside tumor by 3-D PDUS in the present study, performance improved to 89.4% accuracy, 84.9% sensitivity and 93.3% specificity. There were some limitations to this study. There was no intraobserver validation for the same lesion and histopathologic correlation of the vascularity was lacking. In addition, this technique may be difficult to apply if the tumor is associated with prominent posterior shadowing that obscures the tumor vascularity. On the other hand, the proposed method is useful and verified on 113 tumors, including 60 benign and 53 malignant tumors. The dataset might be too small to completely prove the robustness of this diagnosis system. Therefore, a larger cohort study should be collected for robustness verification of this system in the future. In conclusion, separation of the vascular information inside and outside 3-D isosurface tumor contour obtained
1868
Ultrasound in Medicine and Biology
Volume 38, Number 11, 2012
Fig. 9. An example of false positive in a case of fibrocystic changes. The malignancy probability evaluated by the neural out network was 0.95, N0in 5 0, N5in 5 0, VV in 5 0 mm3, N0out 5 10, Bout 5 0, Cout 5 0, Lout all 5 45:54 mm, Lsig 5 18:84 mm, 3 3 out out out 5 0:55, mm VV 5 191.90 mm , MICM 5 3.37, ICM L 5 4.70 and TV 5 542.76 mm . (a) Vessels (orange) Rout mean after thresholding and its thinning skeletons (red and white). (b) Significant vessel trees outside (red) and inside (white) the tumor. (c) The tumor contour (blue) and primary path in the significant vessel tree outside (red) and inside the tumor (white). (d) The original 2-D PDUS.
from 3-D PDUS can provide more delicate quantitative measurement of VM and VT. This approach can significantly improve diagnostic accuracy in breast masses and has potential clinical value. Acknowledgments—The authors thank the National Science Council of the Republic of China for the financial support for this research (Contract No. NSC 96-2221-E-002-268-MY3).
REFERENCES Bullitt E, Ewend MG, Aylward S, Lin WL, Gerig G, Joshi S, Jung I, Muller K, Smith JK. Abnormal vessel tortuosity as a marker of treatment response of malignant gliomas: Preliminary report. Technol Cancer Res Treat 2004;3:577–584. Bullitt E, Gerig G, Pizer SM, Lin W, Aylward SR. Measuring tortuosity of the intracerebral vasculature from MRA images. IEEE Trans Med Imaging 2003;22:1163–1171. Bullitt E, Lin NU, Smith JK, Zeng D, Winer EP, Carey LA, Lin W, Ewend MG. Blood vessel morphologic changes depicted with MR angiography during treatment of brain metastases: A feasibility study. Radiology 2007;245:824–830. Carson PL, Moskalik AP, Govil A, Roubidoux MA, Fowlkes JB, Normolle D, Adler DD, Rubin JM, Helvie M. The 3D and 2D color flow display of breast masses. Ultrasound Med Biol 1997;23: 837–849. Chang RF, Huang SF, Moon WK, Lee YH, Chen DR. Computer algorithm for analysing breast tumor angiogenesis using 3-D power doppler ultrasound. Ultrasound Med Biol 2006;32:1499–1508. Chang RF, Huang SF, Moon WK, Lee YH, Chen DR. Solid breast masses: Neural network analysis of vascular features at three-
dimensional power Doppler US for benign or malignant classification. Radiology 2007;243:56–62. Chen CN, Cheng YM, Lin MT, Hsieh FJ, Lee PH, Chang KJ. Association of color Doppler vascularity index and microvessel density with survival in patients with gastric cancer. Ann Surg 2002;235: 512–518. Cheng WF, Lee CN, Chu JS, Chen CA, Chen TM, Shau WY, Hsieh CY, Hsieh FJ. Vascularity index as a novel parameter for the in vivo assessment of angiogenesis in patients with cervical carcinoma. Cancer 1999;85:651–657. Colleoni M, Rotmensz N, Maisonneuve P, Sonzogni A, Pruneri G, Casadio C, Luini A, Veronesi P, Intra M, Galimberti V, Torrisi R, Andrighetto S, Ghisini R, Goldhirsch A, Viale G. Prognostic role of the extent of peritumoral vascular invasion in operable breast cancer. Ann Oncol 2007;18:1632–1640. Duda RO, Hart PE, Stork DG. Pattern classification. New York: Wiley; 2001. Ferrara KW, Zagar B, Sokil-Melgar J, Algazi VR. High resolution 3D color flow mapping: Applied to the assessment of breast vasculature. Ultrasound Med Biol 1996;22:293–304. Gokalp G, Topal U, Kizilkaya E. Power Doppler sonography: Anything to add to BI-RADS US in solid breast masses? Eur J Radiol 2009;70: 77–85. Gonzalez RC, Woods RE. Digital image processing. Upper Saddle River, N.J: Prentice Hall; 2002. Hagan MT, Demuth HB, Beale MH. Neural network design. Boston: PWS Pub.; 1996. Horowitz E, Sahni S, Mehta DP. Fundamentals of data structures in C11. Summit, NJ: Silicon Press; 2007. Hsiao YH, Kuo SJ, Liang WM, Huang YL, Chen DR. Intra-tumor flow index can predict the malignant potential of breast tumor: Dependent on age and volume. Ultrasound Med Biol 2008;34:88–95.
Vascular morphology and tortuosity analysis d Y.-C. CHANG et al. Huang SF, Chang RF, Moon WK, Lee YH, Chen DR, Suri JS. Analysis of tumor vascularity using three-dimensional power Doppler ultrasound images. IEEE Trans Med Imaging 2008;27:320–330. Huang YL, Kuo SJ, Hsu CC, Tseng HS, Hsiao YH, Chen DR. Computeraided diagnosis for breast tumors by using vascularization of 3-D power Doppler ultrasound. Ultrasound Med Biol 2009;35: 1607–1614. Kimber A, Bowman AW, Cox D, Ross GJS, Gilchrist R, Davison AC, Ohagan A, Lawrance AJ, Pregibon D, Critchley F, Ekholm A, Titterington DM, Williams DA, Copas JB, Palmgren J. Binary regression-models for contaminated data—Discussion. J R Statist Soc Ser B Met 1988;50:253–265. Kohavi R, John GH. Wrappers for feature subset selection. Artif Intell 1997;97:273–324. Krestan CR, Riedl C, Memarsadeghi M, Rudas M, Pfarl G, Helbich TH. 3D-power Doppler ultrasound of breast lesions with a microbubble contrast agent. Acad Radiol 2002;9(Suppl 2):S384–S385. LeCarpentier GL, Roubidoux MA, Fowlkes JB, Krucker JF, Hunt KA, Paramagul C, Johnson TD, Thorson NJ, Engle KD, Carson PL. Suspicious breast lesions: Assessment of 3D doppler US indexes for classification in a test population and fourfold cross-validationscheme. Radiology 2008;249:463–470. Lee JS. Digital image smoothing and the sigma filter. Comput Vis Graph Image Proc 1983;24:255–269. Lee SW, Choi HY, Baek SY, Lim SM. Role of color and power Doppler imaging in differentiating between malignant and benign solid breast masses. J Clin Ultrasound 2002;30:459–464. Malladi R, Sethian JA, Vemuri BC. Shape modeling with front propagation - A level set approach. IEEE Trans Pattern Anal Mach Intell 1995;17:158–175. Meinel LA, Stolpen AH, Berbaum KS, Fajardo LL, Reinhardt JM. Breast MRI lesion classification: Improved performance of human readers with a backpropagation neural network computer-aided diagnosis (CAD) system. J Magn Reson Imaging 2007;25:89–95. Moore DS. The basic practice of statistics. New York: W.H. Freeman and Co.; 2008. Mortensen K, Holck S, Christensen IJ, Skouv J, Hougaard DM, Blom J, Larsson LI. Endothelial cell nitric oxide synthase in peritumoral microvessels is a favorable prognostic indicator in
1869
premenopausal breast cancer patients. Clin Cancer Res 1999;5: 1093–1097. Osher S, Sethian JA. Fronts propagating with curvature-dependent speed—Algorithms based on hamilton-jacobi formulations. J Comput Phys 1988;79:12–49. Palagyi K, Kuba AA. 3D 6-subiteration thinning algorithm for extracting medial lines. Pattern Recogn Lett 1998;19:613–627. Sehgal CM, Arger PH, Rowling SE, Conant EF, Reynolds C, Patton JA. Quantitative vascularity of breast masses by Doppler imaging: Regional variations and diagnostic implications. J Ultrasound Med 2000;19:427–440. Sharma S, Sharma MC, Sarkar C. Morphology of angiogenesis in human cancer: A conceptual overview, histoprognostic perspective and significance of neoangiogenesis. Histopathology 2005;46: 481–489. Song HH. Analysis of correlated ROC areas in diagnostic testing. Biometrics 1997;53:370–382. Strano S, Gombos EC, Friedland O, Mozes M. Color Doppler imaging of fibroadenomas of the breast with histopathologic correlation. J Clin Ultrasound 2004;32:317–322. Suri JS. Advances in diagnostic and therapeutic ultrasound imaging. Boston, London: Artech House; 2008. Toi M, Inada K, Suzuki H, Tominaga T. Tumor angiogenesis in breastcancer—Its importance as a prognostic indicator and the association with vascular endothelial growth-factor expression. Breast Cancer Res Treat 1995;36:193–204. Uzzan B, Nicolas P, Cucherat M. Perret Gr-Y. Microvessel density as a prognostic factor in women with breast cancer: A systematic review of the literature and meta-analysis. Cancer Res 2004;64: 2941–2955. Weidemann A, Stockmann C, Doedens A, Zhang N, Takeda N, Greenberg JI, Cheresh DA, Johnson RS. Deletion of vascular endothelial growth factor in myeloid cells accelerates tumorigenesis. Nature 2008;456:814–818. Yang WT, Tse GMK, Lam PKW, Metreweli C, Chang J. Correlation between color power Doppler sonographic measurement of breast tumor vasculature and immunohistochemical analysis of microvessel density for the quantitation of angiogenesis. J Ultrasound Med 2002;21:1227–1235.