IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 33, NO.
7, JULY 2014 1503
Multi-Dimensional Tumor Detection in Automated Whole Breast Ultrasound Using Topographic Watershed Chung-Ming Lo, Rong-Tai Chen, Yeun-Chung Chang, Ya-Wen Yang, Ming-Jen Hung, Chiun-Sheng Huang*, and Ruey-Feng Chang*, Senior Member, IEEE AbstractAutomated whole breast ultrasound (ABUS) is be- coming a popular screening modality for whole breast examination. Compared to conventional handheld ultrasound, ABUS achieves operator-independent and is feasible for mass screening. How- ever, reviewing hundreds of slices in an ABUS image volume is time-consuming. Acomputer-aideddetection(CADe) systembased on watershed transform was proposed in this study to accelerate the reviewing. The watershed transform was applied to gather similar tissues around local minima to be homogeneous regions. The likelihoods of being tumors of the regions were estimated using the quantitative morphology, intensity, and texture features in the 2-D/3-D false positive reduction (FPR). The collected database comprised 68 benign and 65 malignant tumors. As a result, the proposed system achieved sensitivities of 100% (133/133), 90% (121/133), and 80% (107/133) with FPs/pass of 9.44, 5.42, and 3.33, respectively. The gure of merit of the combination of three feature sets is 0.46whichis signicantlybetter thanthat of other feature sets ( ). In summary, the proposed CADe systembased on the multi-dimensional FPR using the integrated feature set is promisingindetecting tumors inABUSimages. Index TermsAutomated whole breast ultrasound, breast cancer, computer-aided detection, multi-dimensional false positive reduction, watershed segmentation. I. INTRODUCTION B REAST cancer is the most common cancer and is the second leading cause of mortality for women in 2013 [1]. Early detection and treatment of breast cancer have been shown Manuscript received February 17, 2014; revised March 19, 2014; accepted March 28, 2014. Date of publication April 03, 2014; date of current version June 27, 2014. This work was supported in part by the National Science Coun- cilunder Grant NSC 101-2221-E-002-068-MY3, in part by the Ministry of Eco- nomic Affairs under Grant 102-EC-17-A-19-S1-164, in part by the Department of Health under Grant DOH102-TD-C-111-001, and in part by the Ministry of Education (Republic of China) under Grant AE-00-00-06. Asterisk indicates corresponding author. C.-M. Lo, R.-T. Chen, and M.-J. Hung are with the Department of Com- puter Science and Information Engineering, National Taiwan University, Taipei, Taiwan. Y.-C. Chang is with the Department of Medical Imaging, National Taiwan University Hospital and National Taiwan University College of Medicine, Taipei, Taiwan. Y.-W. Yang is with the Department of Surgery, National Taiwan University Hospital and National Taiwan University College of Medicine, Taipei, Taiwan. *C.-S. Huang is with the Department of Surgery, National Taiwan University Hospital and National Taiwan University College of Medicine, Taipei, Taiwan (e-mail: huangcs@ntu.edu.tw). *R.-F. Chang is with Department of Computer Science and Information En- gineering, the Graduate Institute of Biomedical Electronics and Bioinformatics, National Taiwan University, Taipei, Taiwan (e-mail: rfchang@csie.ntu.edu.tw). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TMI.2014.2315206 to be useful in reducing mortality rates [2]. Two screening modalities, mammography [3], [4] and breast ultrasound (US) [5], [6] are the popular modalities for the detection and diag- nosis of breast tumor. Conventionally, mammography is the main imaging tool on clinical examinations. The drawbacks of mammography are high false positive rate, and not sen- sitive for women with dense breast tissue [7][9]. US is an adjunct imaging modality to the mammography for detecting tumors in dense breasts. Compared to mammography, the US examination is radiation-free and makes patients feel less pain [10][12]. The disadvantages of conventional handheld US are operator dependent and poorly reproducible. For these issues, automated whole breast US (ABUS) imaging systems have been developed to automatically scan the whole breast. Using ABUS in conjunction with mammography resulted in signi- cantly increased cancer detection rates [7]. The advantages of ABUS in less operator dependent and greater reproducibility for follow-up studies has also been shown [7], [13]. How- ever, reviewing hundreds of slices in 3-D ABUS images to discover suspicious abnormalities is time-consuming. The computer-aided detection (CADe) systems of ABUS images have been proposed to accelerate the reviewing procedure and reduce oversight errors [14][18]. Ikedo et al. [16] proposed a fully automatic scheme for detecting mass. The Canny edge detector was used to detect edges followed by the classication of near-vertical edges or near-horizontal edges. The near-vertical edges were regarded as the positions of tumor candidates. Then, the watershed transform was adopted to segment the located positions and generated the tumor candidate regions. Chang et al. [15] devel- oped a CADe system for detecting breast tumors in multi-pass automated breast US. After image preprocessing, the gray level slicing method was adopted to segment the tumor candidates. Seven quantitative features were then used as the criteria in distinguishing between tumors and nontumors. Recently, Moon et al. [17] proposed a CADe system based on multi-scale blob analysis. After the speckle noise reduction, Hessian analysis was used to detect the tumor candidates in ABUS images. Then, three categories features including blobness, internal echo, and morphology features were extracted to classify tu- mors and nontumors. Another approach was two-stage CADe suggested by Tan [18]. Coronal speculation patterns, blobness, contrast, and depth were extracted from voxels for evaluating the malignancy of potential abnormalities. The following re- gion segmentation provided quantitative features for the nal classication by neural-network classier. 0278-0062 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. 1504 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 33, NO. 7, JULY 2014 In this study, a CADe systembased on topographic watershed segmentation was implemented to extract potential abnormali- ties in ABUS images. The watershed transform gathered similar tissues around local minima to become homogeneous regions. The likelihoods of being tumors of the regions were estimated using the quantitative morphology, intensity, and texture fea- tures in 2-D/3-Dfalse positive reduction (FPR). The discrimina- tion ability of the different feature sets were evaluated using the free-response operating characteristics (FROC) [19], [20] curve and the jackknife alternative of FROC-1 (JAFROC-1) gure of merit (FOM) [21]. At different sensitivity rates, the false posi- tives (FPs) per pass were estimated for the feasibility on clinical use. II. MATERIALS AND METHODS A. Patients and Data Acquisition ABUS images collected in this study were acquired using an ACUSON S2000 Automated Breast Volume Scanner (Siemens Medical Solutions, Mountain View, CA, USA) between July and December 2012 in the Breast Center of National Taiwan University Hospital. The ABUS scanner equipped a 14L5BV linear array transducer with variable frequencies, ranging from 5 to 15 MHz. In each scanning, the ABUS generated 318 2-D slices with the thickness of 0.5 mm. To completely scan a breast, three scannings including anteriorposterior, medial and lateral passes were performed. A total of six ABUS image volumes were acquired for a patient. Informed consent and the approval for this retrospective study were obtained from our institutional review board. 133 biopsy-proven lesions (size range: 0.37.2 cm, mean: cm) from104 women (age range: 2180 years, mean years) including 68 benign and 65 malignant le- sions were collected. In benign lesions, 36 brocystic changes, 23 broadenomas, and nine papillomas were included. The ma- lignant lesions included 51 invasive ductal carcinoma (IDC) and 14 ductal carcinoma in situ (DCIS). 34 normal passes with no pathology-proven tumors were also included in the experiment for the generalization of the proposed CADe system. The ratio of abnormal to normal passes was a controlled factor. A total of 138 passes with 104 abnormal passes and 34 normal passes composed the image database. B. Watershed Transform for Tissue Segmentation The proposed CADe system based on the analysis of regions composed of various tissues. The toboggan-based watershed [22] was performed on the transverse planes of 2-D slices in- stead of using time-consuming 3-D segmentation for tissue seg- mentation. The acquired ABUS passes were composed of 318 slices. For a better efciency, every ve slices with slight vari- ation in composition were overlapped in the transverse direc- tion to reduce the number of slices. The hypoechogenicity of breast tumor was maintained using the minimum intensity pro- jection [23] in the overlapping of successive 2-D slices. After overlapping, the local minimum regions in a slice were consid- ered to be catchment basins in topography. Conceptually, the toboggan-based watershed can be regarded as a top-down gra- dient descent approach. This approach found a gradient descent Fig. 1. Topographic view of watershed transform. Fig. 2. Watershed transform combined with the watershed depth produced a hierarchy of catchment basins. path from each pixel to the local minimum of the topographic surface. Pixels slid into the same regional minimum were gath- ered together as a catchment basin. The boundaries between the catchment basins were known as watershed lines. The number of segmented regions can be normalized using the watershed depth [24]. The minimum watershed depth was the difference in height between a local minimum and the adjacent lowest boundary points, as shown in Fig. 1. Sequentially combining the catchment basins whose depths fell below the minimum until all the catchment basins were of sufcient depth completed the merging approach. The merging approach was sequential and generated a hierarchy of regions. ALevel value was set as a frac- tion of the maximum watershed depth to preserve meaningful objects and to minimize over-segmentation. As shown in Fig. 2, the minimum Level is 0 and the maximum Level is 1. indicates the initial segmentation while was used in our experiment to achieve the best compromise between con- tour completeness and the number of regions. C. Suspicious Abnormality Extraction Regions with various tissue composition were acquired after watershed transform. An initial cut-off was automatically performed to remove unlikely tumor area according to their echogenicities and anatomical location. Previous study had reported that benign and malignant tumors are hypo-echogenic than other normal breast tissues [25]. To extract the target suspicious abnormalities, regions with hypo-echogenicity compared to other regions were con- sidered. Regions in a slice were sorted according to their average intensity values in ascending order. The criterion for hypo-echogenicity is dened as (1) LO et al.: MULTI-DIMENSIONAL TUMOR DETECTION IN AUTOMATED WHOLE BREAST ULTRASOUND USING TOPOGRAPHIC WATERSHED 1505 Fig. 3. Results of the suspicious abnormality extraction. (a) Original image with a tumor circled. (b) Illustration of the skin, muscle, and shadows in the image. (c) Result after applying watershed transformto (a), and (d) the extracted suspicious abnormalities. In the experiment, equal to or smaller than were extracted for further analysis. That is, 50% of brighter regions were removed according to their hyper-echogenicities on each slice. After removing the hyper-echogenic regions, anatomical lo- cation was the other criterion to determine nontumor regions. A tumor is an abnormal growth of neoplastic tissues existed in the breast area surrounded by skin, muscle, and shadows [Fig. 3(b)]. Removing the regions connected to the image border was ex- pected to be meaningful in decreasing nontumors. As shown in Fig. 3, the leftest and rightest region in (c) corresponds to the shadows in (b). So are the top (skin) and bottom (muscle) re- gions. After removing these border regions, the nal result of the suspicious abnormality extraction is shown in Fig. 3(d). D. 2-D/3-D False Positive Reduction The extracted suspicious abnormalities were the tumor candidates in the further classication. Quantitative features were extracted from the tumor candidates and combined in a classier for the distinguishing between tumors and nontumors. A 2-D/3-D FPR composed of applying 2-D FPR followed by 3-D FPR was used in the classication procedure. In 2-D FPR, the region characteristics were rst described using the quantitative morphology, intensity, and texture features. For each region, the quantitative features were combined to generate a predicted probability being a tumor in a classier. Nontumor regions with lower probabil- ities than true tumor regions were ltered out in 2-D FPR. For 3-D information, the connected component method [26] TABLE I QUANTITATIVE FEATURE CATEGORIES USED IN 2-D/3-D FALSE POSITIVE REDUCTION was used to merge adjacent 2-D regions with overlapping between slices to be a 3-D region. Also, 3-D morphology and intensity features were extracted to remove regions with lower probabilities than true tumor regions in 3-D FPR. Based on this approach, the spatial information of 2-D and 3-D spaces was modeled to lter out the regions with low likelihoods being tumors. The quantitative features are described in the following sections and are summarized in Table I. 1) Morphology Features: Morphology features extracted from tumor contour can provide useful information for tumor detection [17]. Four morphology features including region size, compactness, size to bounding box ratio, and short to long axis ratio were implemented. The delineation of region boundary was obtained after the watershed transform. By calculating the region area [ (2-D)] and region volume [ (3-D)], noises or shadows with extremely small and large sizes can be distinguished and be removed. A tumor is formed by an abnormal growth of neoplastic tissues. In other words, a tumor can be regarded as a cluster of homogeneous region. The region size to bounding box ratio was used to describe the aggregation property of the region. The region volume to 3-D bounding box was also 1506 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 33, NO. 7, JULY 2014 calculated. The 2-D and 3-D size to bounding box ratio were dened by the following equations: (2) (3) where and are the width and height of the 2-D bounding box, respectively. , , and are the width, height, and length of the 3-D bounding box, re- spectively. Fat tissues have at and narrow shapes in ABUS images. Ac- cording to this property, the short to long axis ratio was used to remove the nontumor regions with a fat-like shape. The 2-D and 3-D short to long axis ratio are dened as (4) (5) Another useful shape descriptor of a region is compactness [28]. Compactness, , is used to describe the relation between the contour and the region area. The compactness value is 1 on a regular and round region, while the value is within (0, 1] on an irregular and fragmented object. The compactness formulas for 2-D and 3-D region are dened by (6) (7) where indicates the length of the 2-D regions perimeter and represents the surface area of a 3-D region. 2) Intensity Features: The echogenicity of a tumor is lower than surrounding normal tissues in ABUS images. The intensity mean, , was then used to remove hyper-echogenicities. The 2-D and 3-D intensity mean are dened as follows: (8) (9) where and are the intensity value of pixels (2-D) and voxels (3-D) in a region, respectively. The intensity distribution inside a tumor was also quantied for the uniformity property. Whether a region is composed of identical tissues or is a mixture of various tissues can be ob- served by the intensity standard deviation (SD). The following equations dened the intensity SD of 2-D and 3-D region, re- spectively (10) (11) Fig. 4. Pixel pair of four directions from the centered pixel ( ). Pixel 1 to 4 are the neighboring pixels in the direction of 0 , 45 , 90 , and 135 with , respectively. The intensity rank mentioned previously was also a quantitative feature. The intensity value of a region relative to other regions would reveal the echogenic characteristics. The contrasts between a region and its adjacent neighbors were quantied. Two regions are adjacent if they share a common boundary. The intensity difference of adjacent regions is dened as (12) where is the set of adjacent regions and is the number of the adjacent regions. 3) Texture Features: Texture is one of the important features in identifying an object in an image. Texture features can be classied into two groups according to the statistics: rst-order features and second-order features. The rst-order features extracted from intensity histograms shows the entirety proper- ties. The extended second-order features take the correlations between pixels and their neighbors into consideration. The gray level co-occurrence matrix (GLCM) method [29] which describes the second-order texture features have shown to be useful in pattern recognition applications [30], [31]. To gen- erate GLCM features, an original image was quantied into an image with reduced number of intensity bins, . Afterward, the co-occurrence matrices were computed from by scanning each pixel and its neighboring pixels. Each element is the frequencies of two neighboring pixels separated by a distance and the direction is given by the angle , one with gray value and the other with gray value . and , 45 , 90 , 135 , shown in Fig. 4 was used in the experiment [32]. That is, the GLCM features were extracted from four co-occurrence matrices with different angles. Eight GLCM texture features are dened as follows: (13) (14) LO et al.: MULTI-DIMENSIONAL TUMOR DETECTION IN AUTOMATED WHOLE BREAST ULTRASOUND USING TOPOGRAPHIC WATERSHED 1507 Fig. 5. Series of the results after applying the 3-D FPR. (a) he original ABUS images with a tumor circled in the middle slice. (b) Result after applying watershed transform to (a). (c) Remaining tumor candidates after suspicious abnormality extraction. (d) Detected results after applying 2-D FPR. (e) Final detected results after applying 3-D FPR. Solid circle indicates the tumor area specied in pathology report and the dot circle indicates the false positive area. (15) (16) (17) (18) (19) (20) where , , , and are mean and standard deviation of the marginal distributions of . (21) (22) The classication results of 2-D FPR and 3-D FPR are shown in Fig. 5. The suspicious abnormalities in the individual slice of Fig. 5(c) are removed by 2-D FPR to get Fig. 5(d). The nal classication result presented in Fig. 5(e) is after applying the 3-D FPR to the three slices in Fig. 5(d). 4) Statistical Analysis: The proposed features were tested if they were signicant in distinguishing tumors and nontumors in 2-D and 3-D FPR. The KolmogorovSmirnov test [33] was rst applied to determine whether the feature is a normal distribution or not. If the distribution of a feature was normal, the mean and SD values of the features were calculated to estimate the difference between the tumors and nontumors with Students 1508 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 33, NO. 7, JULY 2014 TABLE II SENSITIVITIES OF THE CADE SYSTEM AND THE CORRESPONDING FPS/PASS USING DIFFERENT FEATURE SETS test [33]. If a feature was nonnormal distribution, the median value was computed and the MannWhitney U test [33] was used. A -value less than 0.05 was considered to be statistical signicance. The quantitative features were used to train the binary lo- gistic regression classier[34] for region classication. The performance was evaluated with ten-fold cross-validation. For each region, the predicted probability as the classication result was between 0 and 1. A region with probability greater than a threshold was regarded as a tumor. Otherwise, it was classied as a nontumor. The ground truth of breast mass was drawn by an expert breast radiologist. The tumor area was checked according to the original US examination report. If the distance between the center of a candidate region and a true tumor is shorter than 2 cm, the candidate region was determined to be a true positive. After manually marking the ground truth of masses, the proposed CADe system can construct a tumor likelihood model automatically. All steps in the CADe system were automatic. The trade-offs between the sensitivity of tumor detection and number of FPs/pass were obtained by different thresholds and were illustrated using FROC curve [19], [20]. The performance of the CADe system was also evaluated using JAFROC-1FOM [21]. The FOM describes an average value that whether the probabilities of tumors were higher than the highest FP probability in a pass. A CADe system with a higher FOM indicates a more reliable suggestion about the likelihood of being a tumor. The FOM is dened as follows: (23) (24) where is the number of passes, is the number of tumors in all passes, is the number of tumors in th pass, is the number of nontumors in th pass, and the function rep- resents the predicted probability of a tumor or a nontumor . In the function , and are the predicted probabilities of a tumor and a nontumor, respectively. If the predicted probability of a tumor is greater than that of a nontumor, the function out- puts a value of 1. JAFROC-1 FOM was used in analyzing the performances of the CADe systems using different feature sets. The evaluated feature sets were composed of one or more fea- ture sets. In the multiple comparisons between feature sets, the BonferroniHolm correction method [35] was used to generate more reliable type I error (FP) rate and -values. III. RESULTS The proposed CADe system was evaluated with an ABUS database including 104 abnormal passes and 34 normal passes. The pathology-proven tumors in abnormal passes were com- posed of 68 benign and 65 malignant tumors. In the result of suspicious abnormality extraction, the maximum, minimum, average and standard deviation of the suspicious abnormality number per pass were 470, 53, 291.61, and 92.90, respectively. Upon the three feature sets, the average number of FPs/pass at the sensitivity of 100% was reduced from 291.61 to 18.19 via the 2-D FPR and further reduced to 9.44 in the following 3-D FPR. For detecting the malignant tumors only, the pro- posed CADe system achieved a sensitivity of 100% with 9.10 FPs/pass. The discrimination abilities of the three feature sets in FPR were shown in Table II. In the statistical analysis, The JAFROC-1 FOMs of different feature sets are listed and are compared in Table III. The FOMs of the morphology, intensity, and texture feature set are 0.14, 0.18, and 0.17, respectively. Upon the complementary advantage, the best performance ( ) was achieved by the combined feature set including morphology, intensity, and texture features. The adjusted -values for FOMs between different combinations of feature sets are also listed. The combination of different feature sets can signicantly improve the FOM in tumor detection. The details of different trade-offs between sensitivity and FPs/pass for different feature sets are illustrated using FROC curves in Fig. 6. Table IV shows the distribution of missed tumors at sensi- tivity from 50% to 90%. The numbers of missed malignant and benign tumors were very close above the sensitivity of 70%. No matter what size a tumor was, the number of missed tumors de- creased with the higher sensitivity. To investigate the generalization of the proposed CADe system, the average FP numbers of the abnormal pass and normal pass at different sensitivities are listed in Table V. The FP number of normal passes was slightly more than that of abnormal passes at all sensitivity rates. All tumors specied in the pathology report were detected by the proposed CADe system. Fig. 7 shows a true positive case of 1.51 cm broadenomas. The case in Fig. 8 illustrates a true positive case of 4.46 cm ductal carcinoma in situ and a FP. In Fig. 9, a rib is shown to be a detected FP. The experiment was accomplished on an Intel Core i7-2600 3.4 GHz processor with LO et al.: MULTI-DIMENSIONAL TUMOR DETECTION IN AUTOMATED WHOLE BREAST ULTRASOUND USING TOPOGRAPHIC WATERSHED 1509 TABLE III ADJUSTED -VALUES OF FOMS BETWEEN DIFFERENT COMBINATIONS OF FEATURE SETS Fig. 6. FROC curve and the corresponding JAFROC-1 FOM for each feature sets. 4 GB RAM. The average running time of the proposed CADe system was 74.3 s/pass. IV. DISCUSSION The motivation of this study was developing a quantitative tumor detection method to assist radiologists in reviewing ABUS images. Based on the target, the sensitivity of 100% should be the performance index for the detection ability. In our result, the performance of combining morphology, inten- sity, and texture feature sets in region classication achieved the sensitivity of 100% (133/133) with 9.44 FPs/pass. The range of FP marks in the collected cases is from 0 to 29. All malignant and benign tumors specied in the pathology report were detected. The acquired ABUS pass was composed of 318 slices. The time-consuming task of reviewing 318 slices may be reduced to observe 9.44 FPs plus 0.96 (133/138) true positives with the assistance of the proposed CADe system. The suggested quantitative features used in FPR were designed for the characteristics owned by both malignant and benign tumors since circumscribed malignant lesions with round shape are often high-grade tumors [36]. A very high percentage of malignant lesions are taller-than-wide. They are less than 7 or 8 mm in greatest diameter. Most invasive cancers larger than 1 cm are round or wider-than-tall. Detecting and providing all tumors in the ABUS images for radiologists to do malignancy evaluation would be better than only the potential malignant tumors suggested by CADe systems. Tan et al. [18] developed a two-stage malignancy detection system for ABUS images. Coronal speculation patterns, blobness, contrast, and depth were extracted from voxels for evaluating the malignancy of potential abnormalities. The following region segmentation provided quantitative features for the nal classication by neural-network classier. The performance achieved the sensi- tivity of 64% with 1 FP/pass which is slightly better than the proposed CADe system (60%, 1.58 FPs). However, 12% of the malignant tumors is missed in the initial stage. The missed malignant tumors may have benign-like shape. In this study, the proposed CADe system achieved the sensitivity of 100% with quantitative morphology, intensity, and texture feature sets. The corresponding 9.44 FPs/pass can be further decreased after adding the landmarks suggested by Tan et al. in the future. The nding of imperfect instances of landmarks within a certain class of shapes may be time-consuming. Moon et al. [17] pro- posed a CADe system based on multi-scale Hessian analysis. The detection rates of ABUS tumors were 100%, 90%, and 70% with 17.4, 8.8, and 2.7 FPs/pass, respectively. The blob detection based on Hessian analysis was used in detecting tumors with low contrast edges or tumors closed to shadows. We also applied the multi-scale Hessian analysis to the same dataset used in this study to provide a performance comparison. The tumor detection rates were 100%, 90%, and 70% with 18.0, 9.1, and 4.3 FPs/pass, respectively. The performance is close to the result in the article but worse than the sensitivity of 100% with 9.44 FPs/pass in the proposed CADe algorithm. In the result observation, many shadows distributed in the image border had blob-like structures and were classied to be FPs. In this study, many shadows were removed in suspicious abnormality extraction as shown in Fig. 3. Additionally, the average running time of multi-scale Hessian analysis was about 13 minutes per pass, the computational processing may be not practical in clinical use. The average running time of the proposed CADe system was 74.3 s/pass. For the possible appli- cation in screening, normal passes with no pathology-proven tumors were also included in the experiment to evaluate the proposed CADe system. The number of FPs/pass is 10.93 for normal passes. The CADe system detected potential carcinomas with hypo- echoic properties which is the most common presentation of in- vasive carcinomas. However, the limitation is that it may not perform well for all kinds of breast carcinomas. Tumors such as lobular carcinoma presenting with little or no mass may not be segmented well. Some mucinous/colloid carcinomas are hyper- echoic or isoechoic which is also a challenge to the detection al- gorithm. Microcalcications along ducts with ductal thickening in rare invasive ductal carcinomas are not considered. To eval- uate the generalization ability, the CADe system will be used in the screening database with hundreds of normal women since 1510 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 33, NO. 7, JULY 2014 TABLE IV MISSED TUMOR NUMBERS FOR DIFFERENT SIZES OF BENIGN AND MALIGNANT TUMORS AT DIFFERENT SENSIVITITY TABLE V SENSITIVITY OF THE CADE SYSTEM AND THE CORRESPONDING FPS/PASS WITH ABNORMAL PASSES, NORMAL PASSES, AND ALL PASSES Fig. 7. True positive case of 1.51 cmbroadenomas. (a) Original ABUS image. (b) Potential tumor regions delineated by watershed segmentation. (c) Detected tumor with the predicted likelihood of 0.53 after 2-D/3-D false positive reduc- tion. Solid circle indicates the tumor area specied in pathology report. invasive cancers occur about 37 per thousand women. Another limitation of this study is that the acquired ABUS image data- base was collected from a single medical center. To evaluate the generalization of the CADe system, ABUS databases from multicenter and different manufactories should be included in the future study. More quantitative features may need to be cal- culated for various databases. For example, combining GLCM features generated by different distance parameters for multi- resolution analysis would be useful to process datasets fromvar- ious ABUS systems. Using more anatomical information existed in the 3-D breast structure would be also useful in estimating likelihoods of being tumors. The FPs of fat tissues, ribs, and shadows could be reduced further. For a better segmentation, intensity inhomogeneity correction [37] can be used in US images to suppress the shadows and avoid boundary leakage. Removing Fig. 8. True positive case of 4.46 cm ductal carcinoma in situ. (a) Original ABUS image. (b) Potential tumor regions delineated by watershed segmenta- tion. (c) Detected tumor with the predicted likelihood of 0.53 after 2-D/3-Dfalse positive reduction. Solid circle indicates the tumor area specied in pathology report. Dot circle indicates the false positive and the predicted likelihood was 0.76. Fig. 9. False positive case. (a) Original ABUS image. (b) Potential tumor re- gions delineated by watershed segmentation. (c) Detected rib region with the predicted likelihood of 0.61 after 2-D/3-D false positive reduction. Dot circle indicates the false positive area. speckle noise in the ABUS images would be also useful for segmentation. Anisotropic diffusion [38] is widely used to reduce speckle noises and preserve edges simultaneously. With LO et al.: MULTI-DIMENSIONAL TUMOR DETECTION IN AUTOMATED WHOLE BREAST ULTRASOUND USING TOPOGRAPHIC WATERSHED 1511 the use of lattice Boltzmann model [39], the additional overload of diffusion partial differential equation can be solved via GPU. The potential benet of GPU would be considered to improve the efciency in the future study. The less time the system spends, the more useful the system is on clinical examination. The quantitative model established in the proposed CADe system was based on the biopsy-proven pathology and US ex- amination report. For the prospective study, whether radiolo- gists can benet from the detection algorithm in the sensitivity of tumor detection is relevant in clinical use. An ABUS image explorer and marking system is going to be constructed for radi- ologists to mark tumors and review the abnormalities suggested by the proposed CADe algorithm. With the quantitative system, recording tumor locations and generating examination report for ABUS database would be more efcient. REFERENCES [1] R. Siegel, D. Naishadham, and A. Jemal, Cancer statistics, 2013, CA, Cancer J. Clinicians, vol. 63, no. 1, pp. 1130, 2013. [2] A. Jemal, F. Bray, M. M. Center, J. Ferlay, E. Ward, and D. Forman, Global cancer statistics, CA, Cancer J. Clinicians, vol. 61, no. 2, pp. 6990, 2011. [3] M. A. Roubidoux, M. A. Helvie, N. E. Lai, and C. Paramagul, Bi- lateral breast cancer: Early detection with mammography, Radiology, vol. 196, no. 2, pp. 427431, 1995. [4] T. E. Wilson, M. A. Helvie, and D. A. August, Breast cancer in the elderly patient: Early detection with mammography, Radiology, vol. 190, no. 1, pp. 203207, 1994. [5] E. A. Sickles, R. Filly, and P. Callen, Benign breast lesions: Ul- trasound detection and diagnosis, Radiology, vol. 151, no. 2, pp. 467470, 1984. [6] T. M. Kolb, J. Lichy, and J. H. Newhouse, Comparison of the perfor- mance of screening mammography, physical examination, and breast US and evaluation of factors that inuence them: An analysis of 27,825 patient evaluations1, Radiology, vol. 225, no. 1, pp. 165175, 2002. [7] K. M. Kelly, J. Dean, W. S. Comulada, and S. J. Lee, Breast cancer detection using automated whole breast ultrasound and mammography in radiographically dense breasts, Eur. Radiol., vol. 20, no. 3, pp. 734742, 2010. [8] N. F. Boyd et al., Heritability of mammographic density, A risk factor for breast cancer, N. Eng. J. Med., vol. 347, no. 12, pp. 886894, 2002. [9] P. Crystal, S. D. Strano, S. Shcharynski, and M. J. Koretz, Using sonography to screen women with mammographically dense breasts, Am. J. Roentgenol., vol. 181, no. 1, pp. 177182, 2003. [10] W. K. Moon, D. Y. Noh, and J. G. Im, Multifocal, multicentric, and contralateral breast cancers: Bilateral whole-breast US in the preoper- ative evaluation of patients1, Radiology, vol. 224, no. 2, pp. 569576, 2002. [11] K. Flobbe, P. Nelemans, A. Kessels, G. Beets, M. Meyenfeldt, and J. Engelshoven, The role of ultrasonography as an adjunct to mammog- raphy in the detection of breast cancer: A systematic review, Eur. J. Cancer, vol. 38, no. 8, pp. 10441050, 2002. [12] K. J. Taylor et al., Ultrasound as a complement to mammography and breast examination to characterize breast masses, Ultrasound Med. Biol., vol. 28, no. 1, pp. 1926, 2002. [13] W. Buchberger, A. Niehoff, P. Obrist, P. DeKoekkoek-Doll, and M. Dnser, Clinically and mammographically occult breast lesions: De- tection and classication with high-resolution sonography, Seminars in Ultrasound, CT MRI, vol. 21, no. 4, pp. 325336, 2000. [14] R. F. Chang, K. C. Chang-Chien, H. J. Chen, D. R. Chen, E. Takada, and W. K. Moon, Whole breast computer-aided screening using free-hand ultrasound, in Int. Congr. Ser., Jun. 2005, vol. 1281, pp. 10751080. [15] R. F. Chang et al., Rapid image stitching and computer-aided detec- tion for multipass automated breast ultrasound, Med. Phys., vol. 37, p. 2063, 2010. [16] Y. Ikedo et al., Development of a fully automatic scheme for detection of masses in whole breast ultrasound images, Med. Phys., vol. 34, p. 4378, 2007. [17] W. K. Moon, Y. W. Shen, M. S. Bae, C. S. Huang, J. H. Chen, and R. F. Chang, Computer-aided tumor detection based on multi-scale blob detection algorithmin automated breast ultrasound images, IEEE Trans. Med. Imag., vol. 32, no. 7, pp. 11911200, Jul. 2013. [18] T. Tan, B. Platel, R. Mus, L. Tabar, R. M. Mann, and N. Karssemeijer, Computer-aided detection of cancer in automated 3-D breast ultra- sound, IEEE Trans. Med. Imag., vol. 32, no. 9, pp. 16981706, Sep. 2013. [19] D. P. Chakraborty, Maximum likelihood analysis of free-response re- ceiver operating characteristic (FROC) data, Med. Phys., vol. 16, pp. 561568, 1989. [20] S. Yu and L. Guan, A CAD system for the automatic detection of clus- tered microcalcications in digitized mammogramlms, IEEE Trans. Med. Imag., vol. 19, no. 2, pp. 115126, Feb. 2000. [21] D. P. Chakraborty, E. S. Breatnach, M. V. Yester, B. Soto, G. T. Barnes, and R. G. Fraser, Digital and conventional chest imaging: A modied ROC study of observer performance using simulated nodules, Radi- ology, vol. 158, no. 1, pp. 3539, 1986. [22] E. N. Mortensen and W. A. Barrett, Toboggan-based intelligent scis- sors with a four-parameter edge model, in Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit., 1999, vol. 2, pp. 452458. [23] C. Beigelman-Aubry, C. Hill, A. Guibal, J. Savatovsky, and P. A. Gre- nier, Multidetector row CT and postprocessing techniques in the as- sessment of diffuse lung disease1, Radiographics, vol. 25, no. 6, pp. 16391652, 2005. [24] A. P. Mangan and R. T. Whitaker, Partitioning 3-D surface meshes using watershed segmentation, IEEE Trans. Visualizat. Comput. Graph., vol. 5, no. 4, pp. 308321, 1999. [25] P. Skaane and K. Engedal, Analysis of sonographic features in the differentiation of broadenoma and invasive ductal carcinoma, AJR. Am. J. Roentgenol., vol. 170, no. 1, pp. 109114, 1998. [26] A. Rosenfeld and J. L. Pfaltz, Sequential operations in digital picture processing, J. ACM, vol. 13, no. 4, pp. 471494, 1966. [27] M. C. Yang, C. S. Huang, J. H. Chen, and R. F. Chang, Whole breast lesion detection using naive Bayes classier for portable ultrasound, Ultrasound Med. Biol., vol. 38, no. 11, pp. 18701880, 2012. [28] E. Bribiesca, An easy measure of compactness for 2-D and 3-D shapes, Pattern Recognit., vol. 41, no. 2, pp. 543554, 2008. [29] R. M. Haralick, K. Shanmugam, and I. H. Dinstein, Textural features for image classication, IEEE Trans. Syst., Man Cybern., no. 6, pp. 610621, 1973. [30] A. Baraldi and F. Parmiggiani, An investigation of the textural char- acteristics associated with gray level cooccurrence matrix statistical parameters, IEEE Trans. Geosci. Remote Sens., vol. 33, no. 2, pp. 293304, Mar. 1995. [31] W. K. Moon, C. M. Lo, C. S. Huang, J. H. Chen, and R. F. Chang, Computer-aided diagnosis based on speckle patterns in ultrasound im- ages, Ultrasound Med. Biol., vol. 38, no. 7, pp. 12511261, 2012. [32] W. K. Moon, C. M. Lo, J. M. Chang, C. S. Huang, J. H. Chen, and R. F. Chang, Computer-aided classication of breast masses using speckle features of automated breast ultrasound images, Med. Phys., vol. 39, no. 10, pp. 64656473, Oct. 2012. [33] A. Field, Discovering Statistics using SPSS. Thousand Oaks, CA: Sage, 2005. [34] D. W. Hosmer and S. Lemeshow, Applied Logistic Regression. New York: Wiley-Interscience, 2000. [35] S. Holm, A simple sequentially rejective multiple test procedure, Scand. J. Stat., vol. 6, pp. 6570, 1979. [36] W. K. Moon, C. M. Lo, J. M. Chang, C. S. Huang, J. H. Chen, and R. F. Chang, Quantitative ultrasound analysis for classication of BI-RADS category 3 breast masses, J. Dig. Imag., vol. 26, no. 6, pp. 10911098, 2013. [37] C. Y. Lee, Y. H. Chou, C. S. Huang, Y. C. Chang, C. M. Tiu, and C. M. Chen, Intensity inhomogeneity correction for the breast sono- gram: Constrained fuzzy cell-based bipartitioning and polynomial sur- face modeling, Med. Phys., vol. 37, no. 11, pp. 56455654, Nov. 2010. [38] Y. J. Yu and S. T. Acton, Speckle reducing anisotropic diffusion, IEEE Trans. Image Process., vol. 11, no. 11, pp. 12601270, Nov. 2002. [39] Y. Zhao, Lattice Boltzmann based PDE solver on the GPU, Vis. Comput., vol. 24, no. 5, pp. 323333, 2008.