Computer-Based Margin Analysis of
Breast Sonography for Differentiating
Malignant and Benign Masses
Chandra M. Sehgal, PhD, Theodore W. Cary, BA, Sarah A. Kangas, BSN, Susan P. Weinstein, MD, Susan M. Schultz, RDMS, Peter H. Arger, MD, Emily F. Conant, MD
Objective.To evaluate the role of quantitative margin features in the computer-aided diagnosis of malignant and benign solid breast masses using sonographic imaging. Methods.Sonographic images from 56 patients with 58 biopsy-proven masses were analyzed quantitatively for the following fea-tures: margin sharpness, margin echogenicity, and angular variation in margin. Of the 58 masses, 38 were benign and 20 were malignant. Each feature was evaluated individually and in combination with the others to determine its association with malignancy. The combination of features yielding the high-est association with malignancy was analyzed by logistic regression to determine the probability of malignancy. The performance of the probability measurements was evaluated by receiver operating characteristic analysis using a round-robin technique. Results.Margin sharpness, margin echogenici-ty, and angular variation in margin were significantly different for the malignant and benign masses (P< .03, 2-tailed Student ttest). According to quantitative measures, tumor-tissue margins of the malignant masses were less distinct than for the benign masses. Although the mean size of the lesions for the two groups was the same, the mean age of the patients was statistically different (P= .000625). After logistic regression analysis, the individual features age, margin sharpness, margin echogenicity, and angular variation in margin were found to be associated with the probability of malignancy (P< .03). The area under the receiver operating characteristic curve ± SD for the 3-feature logistic regression model combining age, margin echogenicity, and angular variation of margin was 0.87 ± 0.05. Conclusions.The proposed quantitative margin features are robust and can reliably measure margin distinctiveness. These features combined with logistic regression analysis can be useful for computer-aided diagnosis of solid breast lesions. Key words:breast imaging; breast neoplasm; breast sonography; computer-aided diagnosis; sonography.
Received February 10, 2004, from the Department of Radiology, University of Pennsylvania, Philadelphia, Pennsylvania USA. Revision requested March 29, 2004. Revised manuscript accepted for publication April 29, 2004.
This research was supported in part by National Institutes of Health grants CA85424 and CA87526. Preliminary work was presented at the Annual Meeting of SPIE—The International Society for Optical Engineering, Medical Imaging, San Diego, California, February 14–19, 2004.
Address correspondence and reprint requests to Chandra M. Sehgal, PhD, Department of Radiology, University of Pennsylvania Medical Center, 1 Silverstein, 3400 Spruce St, Philadelphia, PA 19104 USA.
E-mail: [email protected]. Abbreviations
AVM, angular variance in margin; Az, area under the curve; CAD, computer-aided diagnosis; M-Echo, margin echogenicity; MGL diff, mean gray level difference; M-Sharp, margin sharpness; ROC, receiver operating characteristic
reast cancer is one of the most common cancers in women, accounting for 21% of cancers diag-nosed.1Sonography, now widely used for
investi-gation of breast lesions,2,3can facilitate diagnosis
of simple cysts in the breast with accuracy of 96% to 100%.4When a solid lesion is present, unequivocal
differ-entiation between benign and malignant masses on sonography alone is more difficult. Diverse approaches have been proposed to improve the accuracy of classifica-tion. Stavros et al5identified 18 sonographic features that
had high predictive value (positive or negative) and high specificity for classifying benign and malignant masses. These features were based on qualitative description of the shape, contour, margin, and echogenicity of the lesions. The use of these features achieves sensitivity of 98.4% and a negative predictive value of 99.5%.6
These early findings have been furthered by sev-eral investigators. Jackson7showed that the use of
feature-based sonography improved the perfor-mance of mammography. Lister et al8 observed
that, in the evaluation of clinically benign, discrete, and symptomatic breast masses, feature-based sonography was more accurate and sensitive than mammography. In a study in which 3 readers’ interpretations of sonographic features were found to be well correlated, Arger et al9showed that the
subgroup of features proposed5had high
specifici-ty and a high positive predictive value.
Although these results are encouraging, it is uncertain whether the same performance can be achieved routinely among radiologists with vary-ing levels of expertise in breast sonography. The lack of uniformity and consensus among observers’ use and interpretation of various descriptive terms for malignant and benign masses often results in inconsistent diagnosis. There is an ongoing effort to develop a standard-ized lexicon for describing sonographically iden-tified solid masses.10,11 There are also efforts to
develop a computer-based system for evaluating sonographic images.6,12–19The benefit of the latter,
if successful, is that the computer assessments would be (1) reproducible from one operator to the next, (2) cost-effective, and (3) helpful in pro-viding second opinions on diagnoses.
Much computer-based analysis statistically evaluates image texture or histograms of the images,14–18 whereas clinical assessments are
made by qualitatively evaluating the shape and margin features of the solid masses.2–5,7–11 The
disconnect between the computer and the clini-cal methods often makes it difficult to assess how the various computer features relate to the clini-cal and biologiclini-cal properties of the solid masses. Our long-term goal is to bridge the gap between the two approaches. As a first step, we describe an approach to differentiating between malig-nant and benign breast masses by evaluating the sharpness and the continuity of their margins using 4 quantitative features. These features are extracted from sonograms and then analyzed to determine their ability to distinguish between malignant and benign masses.
Materials and Methods
Cases for analysis were obtained retrospectively from a breast imaging database of patients with solid masses who underwent biopsy over a
24-month period. Fifty-six cases were identified that had sonograms for review that included both radial and antiradial images of adequate tech-nique (ie, no biopsy needle in the image and no outside films included). Because 2 patients had 2 masses each, 58 masses were studied. The stud-ies were approved by the Institutional Review Committee.
Sonographic imaging was performed with a LOGIQ 700 system (GE Healthcare, Milwaukee, MI) and a 10- to 13-MHz transducer. The sono-graphic films were digitized with a Macintosh G4 computer (Apple Computer, Inc, Cupertino, CA) and a scanner (PowerLook 110/UTA; UMAX Technologies, Fremont, CA) at 300-dots-per-inch resolution and 8-bit (256 shades of gray) depth. Each digitized study was assigned a code name after all patient identification was deleted. The multiple frames on each film were spliced and formatted into a QuickTime movie (Apple Computer, Inc) in which each frame represented a single image.
The tumor-tissue margin in each image was analyzed by a custom program developed in our laboratory using the Interactive Data Language platform (Research Systems, Inc, Denver, CO). The analysis consisted of outlining the lesion margin on each frame by a user followed by extraction of lesion features by the computer. Because of shadowing and inadequate backscat-ter from tissue-lesion boundaries, portions of lesion margins are often masked in the images. When these artifacts were present in the images, they were excluded from the analysis by drawing a region of disinterest. The following 3 quantita-tive features were extracted from each image.
Margin Sharpness
The lesion was divided into N sectors through the center of gravity of the user-defined boundary. The computer automatically drew user-defined 5-pixel-wide bands on either side of the bound-ary by morphologic dilation and erosion using a disk-shaped neighborhood of a 5-pixel radius. Both erosion and dilation were used because the shells were constructed on both on the inside and outside of the margin. In each sector, the mean gray levels of pixels in the inner and outer shells were compared. The 2-tailed Student ttest was used for comparison. The sectors with P val-ues below a user-defined threshold were defined as sectors with distinct margins. The user-defined threshold was determined by analysis of
5 sample images from cases not included in this study. The threshold was kept fixed for all the images analyzed. The margin sharpness (M-Sharp) was calculated as n(distinct) × 100/N, where n(distinct) represents the number of sec-tors with a significant difference in gray level at the boundary (P< .001).
Margin Echogenicity
The margin echogenicity (M-Echo) was defined as the difference in brightness at the tumor mar-gin. This feature was determined by measuring the mean gray level difference (MGL diff) of the inside and the outside shells. Only the sectors with statistically distinct margins were consid-ered in the calculation.
Angular Variance in Margin
The angular variance in margin (AVM) quantifies the inhomogeneity in margin brightness with angle. It was determined by measuring mean and SD of the difference in mean gray level of the inside and outside shells in each sector. The AVM was calculated as the ratio SD/mean.
Feature Analysis
Standard descriptive statistics based on the arithmetic mean and SD were derived for the malignant and benign groups. The 2-tailed Student ttest of unequal variance was used to determine the statistical significance of the dif-ference between the 2 groups.
The features showing a significant difference between the benign and malignant groups (P< .05) were fitted to a logistic regression model to calculate probability of malignancy. A 2-step approach was followed.
First, each feature was fitted to the model logit(Pm) = b0+ b1xto determine the association between each variable and incidence of malignan-cy. The logit(Pm) was back-transformed to proba-bility by the formula Pm= 1/(1 + exp[–logit(Pm)]).
P< .05 was used to test the null hypothesis that malignancy is unrelated to the measured feature.
In the second step, multiple logistic regres-sions were used to determine whether the association between the probability of malig-nancy and features increased when a combina-tion of features was used. A 2 × 2 table was constructed for each analysis to determine the predictive accuracy of the logistic model at a fixed cutoff probability of Pm= 0.5.
Receiver Operating Characteristic Analysis
The regression model that yielded the highest predictive accuracy, with each individual fea-ture contributing significantly (P < .05) to the probability of malignancy, was evaluated for diagnostic performance by receiver operating characteristic (ROC) analysis. A round-robin approach, in which N – 1 samples (of the N
samples in the data) were trained to predict the behavior of the remaining Nth sample, was used to assess the discriminating capability of the image features. The process was repeated until each sample was the test case. The area under the curve (Az) was used as a measure of the diagnostic performance.
Results
General Characteristics
Of the 58 masses, 38 (65.5 %) were benign and 20 (34.5%) were malignant. The mean age of the patients and mean size of the lesions for benign, malignant, and all masses are summa-rized in Table 1. The mean size of lesions stud-ied was 1.2 cm2. On average, malignant masses
were larger than benign masses (1.4 versus 1.1 cm2),
but the difference was not statistically signifi-cant (P = 0.28). The mean age of the patient population studied was 52.4 years. On average, patients with malignant masses were older than patients with benign masses (61.7 versus 47.5 years). The difference in age for the two groups was highly significant (P= .000625).
Table 1.Lesion Size and Age of the Patient Population Studied
Lesion size, cm2 Age, y
Cases Mean ± SD Median Range Mean ± SD Median Range
All (N = 58) 1.2 ± 1.0 0.8 0.1–4.1 52.4 ± 14.8 50.8 24.1–86.7 Benign (n = 38) 1.1 ± 1.0 0.7 0.1–4.1 47.5 ± 12.7 48.2 24.1–82.1 Malignant (n = 20) 1.4 ± 1.0 1.3 0.2–3.2 61.7 ± 14.1 60.2 38.4–86.7
Image Characteristics
Figures 1 and 2 are examples of images of 2 patients with benign and malignant masses, respectively. Figures 1A and 2A show the original images; Figures 1B and 2B show the user-drawn boundaries; and Figures 1C and 2C show the superimposed radial sectors in which the mea-surements were made. The graph in Figure 1D shows a plot of Pvalues as a function of angle; 0° is at the 3-o’clock position, and the angle is mea-sured going counterclockwise. The values below the threshold of P= .001 shown by the horizontal dotted line represent sectors with well-delineated tissue-lesion borders. Qualitatively, the margin of the benign mass in Figure 1 is more sharply defined than in the malignant case in Figure 2. This is confirmed quantitatively by the respective graphs in the two figures. For example, only 3 of 72 sectors have undefined borders for
fibroade-noma, compared with nearly half that are unde-fined in the malignant lesion. The computer-measured M-Sharp, M-Echo, and AVM for the 2 cases were 95% versus 52%, 61 versus 9.5, and 0.31 versus 0.36, respectively.
Table 2 summarizes the mean ± SD of the com-puter features measured for the malignant and benign groups. For each of the 3 measured fea-tures, the margins of the benign masses are bet-ter delineated than those of the malignant masses. The difference between the two groups is statistically significant (P< .03).
Logistic Regression and ROC Analyses
Table 3 summarizes the results of logistic regres-sion. Linear logistic regression of the individual features age, M-Sharp, M-Echo, and AVM show that each feature is associated with probability of malignancy (P< .03; Table 3, column 4, rows 3–5).
Figure 1. Biopsy-proved fibroadenoma in a 41-year-old woman. A, Sonogram showing a hypoechoic nodule. B, Sonogram ing the user-defined boundary of the nodule. C, Sonogram showing computer-drawn radial sectors at 5° intervals. D, Graph show-ing the level of statistical significance (P) for MGL diff at the nodule margin in each sector. The dotted horizontal line represents the threshold of P= .001. Only the sectors with Pvalues below the dotted line were considered to have boundary definition.
The predictive accuracy of each variable, when considered individually, ranged between 67.2% and 74.1% (Table 3, column 7, rows 2–5). When age, M-Echo, and AVM were taken together, each feature retained its association with the probabili-ty of malignancy (P < .05; Table 3, column 4, rows 7–9). The 3-feature model improved the pre-dictive accuracy to 82.5% (Table 3, column 7, row 9). Inclusion of M-Sharp did not improve the accuracy (Table 3, column 7, rows 9 and 14).
The ROC curve for the 3-feature model using the round-robin approach is shown in Figure 3. The probability of malignancy (Pm) was determined by back-transforming logit(Pm) by the formula Pm= 1/(1 + exp[–logit(Pm)]). The Az ± SD was 0.87 ± 0.05 with a 95% confidence limit of 0.76 to 0.95.
Discussion
Despite several technical and scientific advance-ments, the error rate for differentiating benign and malignant breast masses continues to be high.19,20 The use of computer-aided diagnosis
(CAD) as a second reviewer of mammograms and sonograms to reduce the error rate has been suggested by many authors.19
Previous investigations on the role of sonogra-phy in CAD have mainly focused on using image texture to differentiate between malignant and benign masses.14–19However, clinical diagnosis is
usually made by evaluating the shape and mar-gin characteristics of the masses in the sono-grams. Stavros et al,5 Jackson,7 Leucht and
Madjar,3 and Arger et al,9 among many other
Figure 2.Biopsy-proven infiltrating lobular carcinoma in a 60-year-old woman. A, Sonogram showing a hypoechoic nodule with a strong posterior shadow. The posterior margin is not visible because of shadowing. B, Sonogram showing the user-defined bound-ary of the nodule. The box represents the area where margin measurements were not made because of shadowing. C, Sonogram showing computer-drawn radial sectors at 5° intervals. D, Graph showing the level of statistical significance (P) for MGL diff at the nodule margin in each sector. The dotted horizontal line represents the threshold of P= .001. Only the sectors with Pvalues below the dotted line were considered to have boundary definition.
clinical investigators, compiled a list of sono-graphic features with high predictive values and specificity for characterizing breast masses. Although the use of several features for CAD is possible, this was an effort to measure one type of visual feature used frequently in clinical stud-ies: the lesion margin characteristics. The clinical evaluation of the lesion margin consists of assessing its demarcation with respect to the sur-rounding tissues3; the margins are often
described as distinct or indistinct. The goal of this study was to design quantitative computer-based features to measure distinctiveness of the tumor-tissue margin and to evaluate their ability for determining the probability of malignancy.
We used 3 different features to measure margin distinctiveness. The feature M-Sharp measured the fraction of the lesion with a well-delineated border; M-Echo measured the strength of the dif-ference in echogenicity between the lesion and the surrounding tissue; and AVM measured the angular variance in echogenicity around the cir-cumference of the lesion.
We first compared the mean values of the fea-tures between the benign and malignant groups
and found that each of the 3 features was statisti-cally different for the 2 groups (Table 2). All 3 fea-tures, M-Sharp, M-Echo, and AVM, were greater for the benign masses than for the malignant ones.
We also observed that the mean age of the benign group was lower than that of the malig-nant group (47.5 versus 61.7 years). These results are consistent with our previous study in which patients with benign masses were significantly younger than those with malignant masses.13
This result indicates that the age of the patients who undergo sonographic examination could also be used with other features to improve mass differentiation.
The next step was to determine whether it was feasible to predict the probability of malignancy (Pm) from the variables age, M-Sharp, M-Echo, and AVM. A logistic regression model was used for this purpose because it describes the relation-ship of several independent variables to a dichotomous dependent variable such as “malig-nant” or “benign.” This model is a nonlinear transformation of the linear regression, in which the dependent variable logit(Pm) is related to the probability Pmand the independent variables (Xi) by the function
Similar to linear regression, logistic regression gives each regressor variable a coefficient, b,
Table 2. Quantitative Margin Characteristics
M-Sharp, % M-Echo, AVM,
Cases Sectors MGL diff SD/MGL diff
Benign 73.7 ± 10.0 22.9 ± 13.2 0.39 ± 0.13 Malignant 66.5 ± 12.0 14.8 ± 7.7 0.27 ± 0.06
P(2-tailed ttest) .027 .0048 .000009 Values are mean ± SD.
Table 3. Summary of Logistic Regression Analysis
Predictive Value at Threshold of Pm= 0.5, %
Features Constant Coefficient P Y(0|0) Y(1|1) Y(0|0) & Y(1|1) Row Column 1 Column 2 Column 3 Column 4 Column 5 Column 6 Column 7 1 Single-parameter model 2 Age –4.89 0.08 .002 89.5 45.0 74.1 3 M-Sharp 3.75 –0.06 .026 89.5 30.0 69.0 4 Margin-Echo 0.62 –0.07 .028 94.7 20.0 69.0 5 AVM 3.24 –12.20 .002 73.7 55.0 67.2 6 3-Parameter Model 0.52 7 Age 0.08 .008 8 M-Echo –0.07 .050 9 AVM –14.10 .005 89.5 70.0 82.8 10 4-Parameter Model 0.84 11 Age 0.08 .008 12 M-Sharp –0.01 .851 13 M-Echo –0.06 .136 14 AVM –13.81 .007 89.5 70.0 82.8
Y(0|0) indicates percent benign (0) cases correctly classified; Y(1|1), percent malignant (1) cases correctly classified; and Y(0|0) & Y(1|1), per-cent benign and malignant cases correctly classified.
which measures the regressor’s independent contribution to the dependent variable with val-ues of 0 for benign and 1 for malignancy.
We first assessed each feature independently. The individual features were fitted to the logistic regression model. The fit of the data to the model was evaluated by the P value as well as by the model’s ability to classify the benign and malig-nant cases correctly. The results in Table 2 show that each variable could classify the cases with 67% to 74% accuracy. P< .05 indicates that the contribution of each parameter in predicting malignancy was significant (Table 3). When we compared combinations of various parameters by the same approach, the 3-parameter model involving age, M-Echo, and AVM yielded the largest percentage of correctly classified cases. The addition of the fourth parameter M-Sharp did not improve the predictive value (Table 3).
Because the 3-parameter model yielded the best results among the models tested, it was further evaluated by ROC analysis. The Azwas 0.874 ± 0.054. At a fixed sensitivity of 95%, appreciable specificity of 63% was achieved. These results compare well with published per-formance.6,13–19,21It is reasonable to anticipate
that addition of these characteristics as inde-pendent features could further improve the CAD performance.
In our previous study, we assessed the perfor-mance of 4 readers in differentiating breast masses using the characteristics of the solid masses outlined by Stavros et al.5The A
zof
indi-viduals ranged from 0.90 to as high as 0.97. The performance of the approach described here (Az = 0.87) has not yet achieved the human expert performance. However, it must be noted that the readers in our previous study were highly trained breast sonologists. It has not yet been shown that the same level of performance can be achieved in general practice. Furthermore, the CAD analy-sis was limited to margins. Shape-based features are also frequently used by clinicians for diagno-sis, and including them in our analysis could fur-ther improve the performance.
Although absolute differentiation between malignant and benign lesions is the goal of many CAD studies, it is also of some interest to deter-mine whether these methods can be used to reduce the number of biopsies. Figure 4 shows an example of how the approach outlined here can be potentially used for biopsy decision. In the example shown, if only the cases with Pm> .24
were recommended for biopsy, then 26 (68%) of 38 benign cases would not have undergone biopsy. This benefit would be at the expense of missing 1 (5%) of 20 malignant cases. Caution, of course, must be exercised in applying the results of this data set to the routine clinical evaluation
Figure 3. Graph showing ROCs of the 3-feature model.
Figure 4. Graph illustrating the use of threshold probability for biopsy decision. The horizontal solid line represents the probability threshold of P= .24.
of breast masses. In this study, the cases were already determined to go for biopsy. In a real clin-ical setting, the case distribution would also include solid masses that may not require biopsy but only routine or short-term follow-up instead. In such situations, CAD could improve decision confidence and perhaps even help a radiologist decide what sort of follow-up to recommend instead of biopsy. A larger study with data from multiple sources and varying experimental con-ditions (including margin threshold) must be further evaluated. Nevertheless, the results of this study are promising. They show that logistic regression analysis of computer-based margin characteristics can be used to diagnose breast masses. Because Doppler sonography can also aid in diagnosis, combining these 2 methods, in addition to developing new features, could fur-ther improve diagnostic performance.
References
1. Hakinson S, Hunter D. Breast cancer. In: Adam HO, Hunter D, Trichopoulos D (eds). Textbook of Cancer Epidemiology. New York, NY: Oxford University Press; 2002:301–339.
2. Madjar H. The Practice of Breast Ultrasound: Techniques, Findings, Differential Diagnosis. New York, NY: Thieme Medial Publishers, Inc; 2000. 3. Leucht D, Madjar H. Teaching Atlas of Breast
Ultrasound. New York, NY: Thieme Medical Publishers, Inc; 1996.
4. Jackson VP. The role of US in breast imaging. Radiology 1990; 177:305–311.
5. Stavros AT, Thickman D, Rapp CL, Dennis MA, Parker SH, Sisney GA. Solid breast nodules: use of sonogra-phy to distinguish between benign and malignant lesions. Radiology 1995; 196:123–134.
6. Chen CM, Chou YH, Han KC, et al. Breast lesions on sonograms: computer-aided diagnosis with nearly setting-independent features and artificial neural net-works. Radiology 2003; 226:504–514.
7. Jackson VP. Management of solid breast nodules: what is the role of sonography? Radiology 1995; 196:14–15.
8. Lister D, Evans AJ, Burrell HC, et al. The accuracy of breast ultrasound in the evaluation of clinically benign discrete, symptomatic breast lumps. Clin Radiol 1998; 53:490–492.
9. Arger PH, Sehgal CM, Conant EF, et al. Interreader variability and predictive value of US descriptions of solid breast masses: pilot study. Acad Radiol 2001; 8:335–342.
10. Baker JA, Kornguth PJ, Soo MS, Walsh R, Mengoni P. Sonography of solid breast lesions: observer variabili-ty of lesion description and assessment. AJR Am J Roentgenol 1999; 172:1621–1625.
11. Merritt CRB. BIRADS lexicon for breast ultrasound. In: Goldberg BB (ed). The Leading Edge in Diagnostic Ultrasound. Atlantic City, NJ: Thomas Jefferson University; 2001:73–76.
12. Richter K, Heywang-Kobrunner SH, Winzer KJ, et al. Detection of malignant and benign breast lesions with an automated US system: results in 120 cases. Radiology 1997; 205:823–830.
13. Sehgal CM, Arger PH, Rowling SE, Conant EF, Reynolds C, Patton JA. Quantitative vascularity of breast masses by Doppler imaging: regional varia-tions and diagnostic implicavaria-tions. J Ultrasound Med 2000; 19:427–442.
14. Goldberg V, Manduca A, Ewert DL, Gisvold JJ, Greenleaf JF. Improvement in specificity of ultra-sonography for diagnosis of breast tumors by means of artificial intelligence. Med Phys 1992; 19:1475– 1481.
15. Garra BS, Krasner BH, Horii SC, Ascher S, Mun SK, Zeman RK. Improving the distinction between benign and malignant breast lesions: the value of sono-graphic texture analysis. Ultrason Imaging 1993; 15:267–285.
16. Dumane VA, Shankar PM, Piccoli CW, Reid JM, Forsberg F, Goldberg BB. Computer-aided classifica-tion of masses in ultrasonic mammography. Med Phys 2002; 29:1968–1973.
17. Shankar PM, Dumane VA, Piccoli CW, Reid JM, Forsberg F, Goldberg BB. Classification of breast masses in ultrasonic B-mode images using a com-pounding technique in the Nakagami distribution domain. Ultrasound Med Biol 2002; 28:1295–1300. 18. Drukker K, Giger ML, Horsch K, Kupinski MA, Vyborny CJ, Mendelson EB. Computerized lesion detection on breast ultrasound. Med Phys 2002; 29:1438–1446.
19. Chen DR, Chang RF, Kuo WJ, Chen MC, Huang YL. Diagnosis of breast tumors with sonographic texture analysis using wavelet transform and neural net-works. Ultrasound Med Biol 2002; 28:1301–1310.
20. Bird RE, Wallace TW, Yankaskas BC. Analysis of can-cers missed at screening mammography. Radiology 1992; 184:613–617.
21. Hashimoto H, Suzuki M, Oshida M, et al. Quantitative ultrasound as a predictor of node metastases and prognosis in patients with breast cancer. Breast Cancer 2000; 7:241–246.